This PDF file contains the front matter associated with SPIE Proceedings Volume 6914, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and the Conference Committee listing.

The level set framework has proven well suited to medical image segmentation^1-6 thanks to its ability of balancing the contribution of image data and prior knowledge in a principled, flexible and transparent way. It consists of evolving a curve toward the target object boundaries. The curve evolution equation is sought following the optimization of a cost functional containing two types of terms: data terms, which measure the fidelity of segmentation to image intensities, and prior terms, which traduce learned prior knowledge. Without priors many algorithms are likely to fail due to high noise, low contrast and data incompleteness. Different priors have been investigated such as shape¹ and appearance priors.⁷ In this study, we propose a simple type of priors: the area prior. This prior embeds knowledge of an approximate object area and has two positive effects. First, It speeds up significantly the evolution when the curve is far from the target object boundaries. Second, it slows down the evolution when the curve is close to the target. Consequently, it reinforces curve stability at the desired boundaries when dealing with low contrast intensity edges. The algorithm is validated with several experiments using Magnetic Resonance (MR) images and Computed Tomography (CT) images. A comparison with another level set method illustrates the positive effects of the area prior.

A novel physical analogy and associated algorithm is proposed for image segmentation. The physical analogy is that of an electrical network. The network is composed of discrete resistive elements that are interconnected, according to adjacency relations in the image. The resistances vary linearly with the image gradient and exponentially with the voltage across the elements. Segmentation is obtained by applying a voltage between a voltage source and ground that are located at points on the inside and outside of an object, respectively. Those points are identified in an interactive manner. Provided that a sufficiently high voltage is applied, a dichotomization of the voltages in the network occurs. In practice, the solution can be obtained in an iterative manner by gradually increasing the applied voltage. At each step as the source voltage is increased, the resistances are held constant and the Kirchoff's-law system of equations is solved algebraically. It is not clear yet whether image segmentation obtained by this algorithm is globally optimal with respect to a cost-function, as is the case for minimum s-t cut image segmentation. However, like minimum s-t cut algorithm, the resistive-network algorithm does tend to favor segmentation boundaries with shorter, and thus smoother, boundaries. A potential advantage of the resistive-network algorithm is that it is amenable to parallelization. The algorithm was found to provide reasonable results for segmentation of a synthetic image of step boundary degraded by blurring and Gaussian noise, low-contrast computed tomography (CT) of a ball, and CT of a liver lesion.

In this paper, we present a new distribution metric for image segmentation that arises as a result in prediction theory. Forming a natural geodesic, our metric quantifies "distance" for two density functionals as the standard deviation of the difference between logarithms of those distributions. Using level set methods, we incorporate an energy model based on the metric into the Geometric Active Contour framework. Moreover, we briefly provide a theoretical comparison between the popular Fisher Information metric, from which the Bhattacharyya distance originates, with the newly proposed similarity metric. In doing so, we demonstrate that segmentation results are directly impacted by the type of metric used. Specifically, we qualitatively compare the Bhattacharyya distance and our algorithm on the Kaposi Sarcoma, a pathology that infects the skin. We also demonstrate the algorithm on several challenging medical images, which further ensure the viability of the metric in the context of image segmentation.

This paper presents a novel analytic technique to perform shape-driven segmentation. In our approach, shapes are represented using binary maps, and linear PCA is utilized to provide shape priors for segmentation. Intensity based probability distributions are then employed to convert a given test volume into a binary map representation, and a novel energy functional is proposed whose minimum can be analytically computed to obtain the desired segmentation in the shape space. We compare the proposed method with the log-likelihood based energy to elucidate some key differences. Our algorithm is applied to the segmentation of brain caudate nucleus and hippocampus from MRI data, which is of interest in the study of schizophrenia and Alzheimer's disease. Our validation (we compute the Hausdorff distance and the DICE coefficient between the automatic segmentation and ground-truth) shows that the proposed algorithm is very fast, requires no initialization and outperforms the log-likelihood based energy.

Accurate image registration is a necessary prerequisite for many diagnostic and therapy planning procedures where complementary information from different images has to be combined. The design of robust and reliable non-parametric registration schemes is currently a very active research area. Modern approaches combine the pure registration scheme with other image processing routines such that both ingredients may benefit from each other. One of the new approaches is the combination of segmentation and registration ("segistration"). Here, the segmentation part guides the registration to its desired configuration, whereas on the other hand the registration leads to an automatic segmentation. By joining these image processing methods it is possible to overcome some of the pitfalls of the individual methods. Here, we focus on the benefits for the registration task. In the current work, we present a novel unified framework for non-parametric registration combined with energy-based segmentation through active contours. In the literature, one may find various ways to combine these image processing routines. Here, we present the most promising approaches within the general framework. It is based on a single variational formulation of both the registration and the segmentation part. The performance tests are carried out for magnetic resonance (MR) images of the brain, and they demonstrate the potential of the proposed methods.

Atlas-based segmentation is a popular generic technique for automated delineation of structures in volumetric data sets. Several studies have shown that multi-atlas based segmentation methods outperform schemes that use only a single atlas, but running multiple registrations on large volumetric data is too time-consuming for routine clinical use. We propose a generally applicable adaptive local multi-atlas segmentation method (ALMAS) that locally decides how many and which atlases are needed to segment a target image. Only the selected parts of atlases are registered. The method is iterative and automatically stops when no further improvement is expected. ALMAS was applied to segmentation of the heart on chest CT scans and compared to three existing atlas-based methods. It performed significantly better than single-atlas methods and as good as multi-atlas methods at a much lower computational cost.

We propose a new framework to propagate the labels in a heart atlas to the cardiac MRI images for ventricle segmentations based on image registrations. The method employs the anatomical information from the atlas as priors to constrain the initialisation between the atlas and the MRI images using region based registrations. After the initialisation which minimises the possibility of local misalignments, a fluid registration is applied to fine-tune the labelling in the atlas to the detail in the MRI images. The heart shape from the atlas does not have to be representative of that of the segmented MRI images in terms of morphological variations of the heart in this framework. In the experiments, a cadaver heart atlas and a normal heart atlas were used to register to in-vivo data for ventricle segmentation propagations. The results have shown that the segmentations based on the proposed method are visually acceptable, accurate (surface distance against manual segmentations is 1.0 ± 1.0 mm in healthy volunteer data, and 1.6 ± 1.8 mm in patient data), and reproducible (0.7 ± 1.0 mm) for in-vivo cardiac MRI images. The experiments also show that the new initialisation method can correct the local misalignments and help to avoid producing unrealistic deformations in the nonrigid registrations with 21% quantitative improvement of the segmentation accuracy.

We present a new standard atlas of the human brain based on magnetic resonance images. The atlas was generated using unbiased population registration from high-resolution images obtained by multichannel-coil acquisition at 3T in a group of 24 normal subjects. The final atlas comprises three anatomical channels (T₁-weighted, early and late spin echo), three diffusion-related channels (fractional anisotropy, mean diffusivity, diffusion-weighted image), and three tissue probability maps (CSF, gray matter, white matter). The atlas is dynamic in that it is implicitly represented by nonrigid transformations between the 24 subject images, as well as distortion-correction alignments between the image channels in each subject. The atlas can, therefore, be generated at essentially arbitrary image resolutions and orientations (e.g., AC/PC aligned), without compounding interpolation artifacts. We demonstrate in this paper two different applications of the atlas: (a) region definition by label propagation in a fiber tracking study is enabled by the increased sharpness of our atlas compared with other available atlases, and (b) spatial normalization is enabled by its average shape property. In summary, our atlas has unique features and will be made available to the scientific community as a resource and reference system for future imaging-based studies of the human brain.

Whole-brain morphometry denotes a group of methods with the aim of relating clinical and cognitive measurements to regions of the brain. Typically, such methods require the statistical analysis of a data set with many variables (voxels and exogenous variables) paired with few observations (subjects). A common approach to this ill-posed problem is to analyze each spatial variable separately, dividing the analysis into manageable subproblems. A disadvantage of this method is that the correlation structure of the spatial variables is not taken into account. This paper investigates the use of ridge regression to address this issue, allowing for a gradual introduction of correlation information into the model. We make the connections between ridge regression and voxel-wise procedures explicit and discuss relations to other statistical methods. Results are given on an in-vivo data set of deformation based morphometry from a study of cognitive decline in an elderly population.

Purpose: Voxel based morphometry (VBM) is increasingly being used to detect diffusion tensor (DT) image abnormalities in patients for different pathologies. An important requisite for these VBM studies is the use of a high-dimensional, non-rigid coregistration technique, which is able to align both the spatial and the orientational information. Recent studies furthermore indicate that high-dimensional DT information should be included during coregistration for an optimal alignment. In this context, a population based DTI atlas is created that preserves the orientational DT information robustly and contains a minimal bias towards any specific individual data set. Methods: A ground truth evaluation method is developed using a single subject DT image that is deformed with 20 deformation fields. Thereafter, an atlas is constructed based on these 20 resulting images. Thereby, the non-rigid coregistration algorithm is based on a viscous fluid model and on mutual information. The fractional anisotropy (FA) maps as well as the DT elements are used as DT image information during the coregistration algorithm, in order to minimize the orientational alignment inaccuracies. Results: The population based DT atlas is compared with the ground truth image using accuracy and precision measures of spatial and orientational dependent metrics. Results indicate that the population based atlas preserves the orientational information in a robust way. Conclusion: A subject independent population based DT atlas is constructed and evaluated with a ground truth method. This atlas contains all available orientational information and can be used in future VBM studies as a reference system.

The topic of studying the growth of human brain development has become of increasing interest in the neuroimaging community. Cross-sectional studies may allow comparisons between means of different age groups, but they do not provide a growth model that integrates the continuum of time, nor do they present any information about how individuals/population change over time. Longitudinal data analysis method arises as a strong tool to address these questions. In this paper, we use longitudinal analysis methods to study tissue development in early brain growth. A novel approach of multivariate longitudinal analysis is applied to study the associations between the growth of different brain tissues. In this paper, we present the methodologies to statistically study scalar (univariate) and vector (multivariate) longitudinal data, and demonstrate exploratory results in a neuroimaging study of early brain tissue development. We obtained growth curves as a quadratic function of time for all three tissues. The quadratic terms were tested to be statistically significant, showing that there was indeed a quadratic growth of tissues in early brain development. Moreover, our result shows that there is a positive correlation between repeated measurements of any single tissue, and among those of different tissues. Our approach is generic in natural and thus can be applied to any longitudinal data with multiple outcomes, even brain structures. Also, our joint mixed model is flexible enough to allow incomplete and unbalanced data, i.e. subjects do not need to have the same number of measurements, or be measured at the exact time points.

Single Plane Illumination Microscopy (SPIM; Huisken et al., Nature 305(5686):1007-1009, 2004) is an emerging microscopic technique that enables live imaging of large biological specimens in their entirety. By imaging the living biological sample from multiple angles SPIM has the potential to achieve isotropic resolution throughout even relatively large biological specimens. For every angle, however, only a relatively shallow section of the specimen is imaged with high resolution, whereas deeper regions appear increasingly blurred. In order to produce a single, uniformly high resolution image, we propose here an image mosaicing algorithm that combines state of the art groupwise image registration for alignment with content-based image fusion to prevent degrading of the fused image due to regional blurring of the input images. For the registration stage, we introduce an application-specific groupwise transformation model that incorporates per-image as well as groupwise transformation parameters. We also propose a new fusion algorithm based on Gaussian filters, which is substantially faster than fusion based on local image entropy. We demonstrate the performance of our mosaicing method on data acquired from living embryos of the fruit fly, Drosophila, using four and eight angle acquisitions.

The introduction of 2D array ultrasound transducers enables the instantaneous acquisition of ultrasound volumes in the clinical practice. The next step coming along is the combination of several scans to create compounded volumes that provide an extended field-of-view, so called mosaics. The correct alignment of multiple images, which is a complex task, forms the basis of mosaicing. Especially the simultaneous intensity-based registration has many properties making it a good choice for ultrasound mosaicing in comparison to the pairwise one. Fundamental for each registration approach is a suitable similarity measure. So far, only standard measures like SSD, NNC, CR, and MI were used for mosaicing, which implicitly assume an additive Gaussian distributed noise. For ultrasound images, which are degraded by speckle patterns, alternative noise models based on multiplicative Rayleigh distributed noise were proposed in the field of motion estimation. Setting these models into the maximum likelihood estimation framework, which enables the mathematical modeling of the registration process, led us to ultrasound specific bivariate similarity measures. Subsequently, we used an extension of the maximum likelihood estimation framework, which we developed in a previous work, to also derive multivariate measures. They allow us to perform ultrasound specific simultaneous registration for mosaicing. These measures have a higher potential than afore mentioned standard measures since they are specifically designed to cope with problems arising from the inherent contamination of ultrasound images by speckle patterns. The results of the experiments that we conducted on a typical mosaicing scenario with only partly overlapping images confirm this assumption.

This study investigated the feasibility of automatic image registration of MR high-spatial resolution proximal femur trabecular bone images as well as the effects of gray-level interpolation and volume of interest (VOI) misalignment on MR-derived trabecular bone structure parameters. For six subjects, a baseline scan and a follow-up scan of the proximal femur were acquired on the same day. An automatic image registration technique, based on mutual information, utilized a baseline and a follow-up scan to compute transform parameters that aligned the two images. These parameters were subsequently used to transform the follow-up image with three different gray-level interpolators. Nearest neighbor interpolation and b-spline approximation did not significantly alter bone parameters, while linear interpolation significantly modified bone parameters (p<0.01). Improvement in image alignment due to the automatic registration was determined by visually inspecting difference images and 3D renderings. This work demonstrates the first application of automatic registration, without prior segmentation, of high-spatial resolution trabecular bone MR images of the proximal femur. Additionally, effects due to imprecise analysis volume alignment are investigated. Inherent heterogeneity in trabecular bone structure and imprecise positioning of the VOI along the slice (A/P) direction resulted in significant changes in bone parameters (p<0.01). Results suggest that automatic mutual information registration using nearest-neighbor gray-level interpolation to transform the final image ensures VOI alignment between baseline and follow-up images and does not compromise the integrity of MR-derived trabecular bone parameters.

The Iterative Closest Point (ICP) algorithm is an efficient and popular technique for surface registration. It however suffers from the well-known problem of local minima that make the algorithm stop before it reaches the desired global solution. ICP can be improved by the use of landmarks or features. We recently developed a level set capable of evolving on the surface of an object represented by a triangular mesh. This level set permits the segmentation of portions of a surface based on curvature features. The boundary of a segmented portion forms a ridgeline/crestline. We show that the ridgelines/crestlines and corresponding enclosed surfaces extracted by the algorithm can substantially improve ICP registration. We compared the performance of an ICP algorithm in three setups: 1) ICP without landmarks. 2) ICP using ridgelines. 3) ICP using ridgelines and corresponding enclosed surfaces. Our material consists of vertebral body surfaces extracted for a study about the progression of Ankylosing Spondylitis. Same vertebrae scanned at intervals of one or two years were rigidly registered. Vertebral body rims and the end plate surfaces they enclose were used as landmarks. The performance measure was the mean error distance between the registered surfaces. From the one hundred registrations that we performed the average mean error was respectively 0.503mm, 0.335mm and 0.254mm for the three setups. Setup 3 almost halved the average error of setup 1. Moreover the error range is dramatically reduced from [0.0985, 2.19]mm to just [0.0865, 0.532]mm, making the algorithm very robust.

In this paper we present a new method to track bone movements in stereoscopic X-ray image series of the knee joint. The method is based on two different X-ray image sets: a rotational series of acquisitions of the still subject knee that allows the tomographic reconstruction of the three-dimensional volume (model), and a stereoscopic image series of orthogonal projections as the subject performs movements. Tracking the movements of bones throughout the stereoscopic image series means to determine, for each frame, the best pose of every moving element (bone) previously identified in the 3D reconstructed model. The quality of a pose is reflected in the similarity between its theoretical projections and the actual radiographs. We use direct Fourier reconstruction to approximate the three-dimensional volume of the knee joint. Then, to avoid the expensive computation of digitally rendered radiographs (DRR) for pose recovery, we develop a corollary to the 3-dimensional central-slice theorem and reformulate the tracking problem in the Fourier domain. Under the hypothesis of parallel X-ray beams, the heavy 2D-to-3D registration of projections in the signal domain is replaced by efficient slice-to-volume registration in the Fourier domain. Focusing on rotational movements, the translation-relevant phase information can be discarded and we only consider scalar Fourier amplitudes. The core of our motion tracking algorithm can be implemented as a classical frame-wise slice-to-volume registration task. Results on both synthetic and real images confirm the validity of our approach.

Measurement of the cortical thickness from 3D Magnetic Resonance Imaging (MRI) can aid diagnosis and longitudinal studies of a wide range of neurodegenerative diseases. We estimate the cortical thickness using a Laplacian approach whereby equipotentials analogous to layers of tissue are computed. The thickness is then obtained using an Eulerian approach where partial differential equations (PDE) are solved, avoiding the explicit tracing of trajectories along the streamlines gradient. This method has the advantage of being relatively fast and insure unique correspondence points between the inner and outer boundaries of the cortex. The original method is challenged when the thickness of the cortex is of the same order of magnitude as the image resolution since partial volume (PV) effect is not taken into account at the gray matter (GM) boundaries. We propose a novel way to take into account PV which improves substantially accuracy and robustness. We model PV by computing a mixture of pure Gaussian probability distributions and use this estimate to initialize the cortical thickness estimation. On synthetic phantoms experiments, the errors were divided by three while reproducibility was improved when the same patients was scanned three consecutive times.

The motivation of this work is to register MR brain tumor images with a brain atlas. Such a registration method can make possible the pooling of data from different brain tumor patients into a common stereotaxic space, thereby enabling the construction of statistical brain tumor atlases. Moreover, it allows the mapping of neuroanatomical brain atlases into the patient's space, for segmenting brains and thus facilitating surgical or radiotherapy treatment planning. However, the methods developed for registration of normal brain images are not directly applicable to the registration of a normal atlas with a tumor-bearing image, due to substantial dissimilarity and lack of equivalent image content between the two images, as well as severe deformation or shift of anatomical structures around the tumor. Accordingly, a model that can simulate brain tissue death and deformation induced by the tumor is considered to facilitate the registration. Such tumor growth simulation models are usually initialized by placing a small seed in the normal atlas. The shape, size and location of the initial seed are critical for achieving topological equivalence between the atlas and patient's images. In this study, we focus on the automatic estimation of these parameters, pertaining to tumor simulation. In particular, we propose an objective function reflecting feature-based similarity and elastic stretching energy and optimize it with APPSPACK (Asynchronous Parallel Pattern Search), for achieving significant reduction of the computational cost. The results indicate that the registration accuracy is high in areas around the tumor, as well as in the healthy portion of the brain.

A comprehensive framework is proposed for the spatial normalization of diffusion tensor (DT) brain images using tensor-derived tissue attributes. In this framework, the brain tissues are first classified into three categories: the white matter (WM), the gray matter (GM), and the cerebral-spinal fluid (CSF) using the anisotropy and diffusivity information derived from the full tensor. The tissue attributes obtained from this anisotropic segmentation are then incorporated into a very-high-dimensional elastic registration method to produce a spatial deformation field. Finally, the rotational component in the deformation field, together with the estimated underlying fiber direction, is used to determine an appropriate tensor reorientation. This framework has been assessed quantitatively and qualitatively based on a sequence of experiments. A simulated experiment has been performed to evaluate the accuracy of the spatial warping by examining the variation between deformation fields. To verify the tensor reorientation, especially, in the anisotropic microstructures of WM fiber tissues, an experiment has been designed to compare the fiber tracts generated from the DT template and the normalized DT subjects in some regions of interest (ROIs). Finally, this method has been applied to spatially normalize 31 subjects to a common space, the case in which there exist large deformations between subjects and the existing approaches are normally difficult to achieve satisfactory results. The average across the individual normalized DT images shows a significant improvement in signal-to-noise ratio (SNR).

Introduction - Fusion of histology and MRI is frequently demanded in biomedical research to study in vitro tissue properties in an in vivo reference space. Distortions and artifacts caused by cutting and staining of histological slices as well as differences in spatial resolution make even the rigid fusion a difficult task. State-of- the-art methods start with a mono-modal restacking yielding a histological pseudo-3D volume. The 3D information of the MRI reference is considered subsequently. However, consistency of the histology volume and consistency due to the corresponding MRI seem to be diametral goals. Therefore, we propose a novel fusion framework optimizing histology/histology and histology/MRI consistency at the same time finding a balance between both goals. Method - Direct slice-to-slice correspondence even in irregularly-spaced cutting sequences is achieved by registration-based interpolation of the MRI. Introducing a weighted multi-image mutual information metric (WI), adjacent histology and corresponding MRI are taken into account at the same time. Therefore, the reconstruction of the histological volume as well as the fusion with the MRI is done in a single step. Results - Based on two data sets with more than 110 single registrations in all, the results are evaluated quantitatively based on Tanimoto overlap measures and qualitatively showing the fused volumes. In comparison to other multi-image metrics, the reconstruction based on WI is significantly improved. We evaluated different parameter settings with emphasis on the weighting term steering the balance between intra- and inter-modality consistency.

EM and level set algorithms are competing methods for segmenting MRI brain images. This paper presents a fair comparison of the two techniques using the Montreal Neurological Institute's software phantom. There are many flavors of level set algorithms for segmentation into multiple regions (multi-phase algorithms, multi-layer algorithms). The specific algorithm evaluated by us is a variant of the multi-layer level set algorithm. It uses a single level set function for segmenting the image into multiple classes and can be run to completion without restarting. The EM-based algorithm is standard. Both algorithms have the capacity to model a variable number of partial volume classes as well as image inhomogeneity (bias field). Our evaluation consists of systematically changing the number of partial volume classes, additive image noise, and regularization parameters. The results suggest that the performances of both algorithms are comparable across noise, number of partial volume classes, and regularization. The segmentation errors of both algorithms are around 5 - 10% for cerebrospinal fluid, gray and white matter. The level set algorithm appears to have a slight advantage for gray matter segmentation. This may be beneficial in studying certain brain diseases (Multiple Sclerosis or Alzheimer's disease) where small changes in gray matter volume are significant.

This paper presents a novel algorithm of 3D human brain tissue segmentation and classification in magnetic resonance image (MRI) based on region restricted EM algorithm (RREM). The RREM is a level set segmentation method while the evolution of the contours was driven by the force field composed by the probability density functions of the Gaussian models. Each tissue is modeled by one or more Gaussian models restricted by free shaped contour so that the Gaussian models are adaptive to the local intensities. The RREM is guaranteed to be convergency and achieving the local minimum. The segmentation avoids to be trapped in the local minimum by the split and merge operation. A fuzzy rule based classifier finally groups the regions belonging to the same tissue and forms the segmented 3D image of white matter (WM) and gray matter (GM) which are of major interest in numerous applications. The presented method can be extended to segment brain images with tumor or the images having part of the brain removed with the adjusted classifier.

In cochlear implant surgery, an electrode array is permanently implanted in the cochlea to stimulate the auditory nerve and allow deaf people to hear. A minimally invasive surgical technique has recently been proposed--percutaneous cochlear access--in which a single hole is drilled from the skull surface to the cochlea. For the method to be feasible, a safe and effective drilling trajectory must be determined using a pre-operative CT. Segmentation of the structures of the ear would improve trajectory planning safety and efficiency and enable the possibility of automated planning. Two important structures of the ear, the facial nerve and chorda tympani, present difficulties in intensity based segmentation due to their diameter (as small as 1.0 and 0.4 mm) and adjacent inter-patient variable structures of similar intensity in CT imagery. A multipart, model-based segmentation algorithm is presented in this paper that accomplishes automatic segmentation of the facial nerve and chorda tympani. Segmentation results are presented for 14 test ears and are compared to manually segmented surfaces. The results show that mean error in structure wall localization is 0.2 and 0.3 mm for the facial nerve and chorda, proving the method we propose is robust and accurate.

This paper presents a framework for brain classification based on multi-parametric medical images. This method takes advantage of multi-parametric imaging to provide a set of discriminative features for classifier construction by using a regional feature extraction method which takes into account joint correlations among different image parameters; in the experiments herein, MRI and PET images of the brain are used. Support vector machine classifiers are then trained based on the most discriminative features selected from the feature set. To facilitate robust classification and optimal selection of parameters involved in classification, in view of the well-known "curse of dimensionality", base classifiers are constructed in a bagging (bootstrap aggregating) framework for building an ensemble classifier and the classification parameters of these base classifiers are optimized by means of maximizing the area under the ROC (receiver operating characteristic) curve estimated from their prediction performance on left-out samples of bootstrap sampling. This classification system is tested on a sex classification problem, where it yields over 90% classification rates for unseen subjects. The proposed classification method is also compared with other commonly used classification algorithms, with favorable results. These results illustrate that the methods built upon information jointly extracted from multi-parametric images have the potential to perform individual classification with high sensitivity and specificity.

Wireless Capsule Endoscopy (WCE) is a colour imaging technology that enables detailed examination of the interior of the gastrointestinal tract. A typical WCE examination takes ~ 8 hours and captures ~ 40,000 useful images. After the examination, the images are viewed as a video sequence, which generally takes a clinician over an hour to analyse. The manufacturers of the WCE provide certain automatic image analysis functions e.g. Given Imaging offers in their Rapid Reader software: The Suspected Blood Indicator (SBI), which is designed to report the location in the video of areas of active bleeding. However, this tool has been reported to have insufficient specificity and sensitivity. Therefore it does not free the specialist from reviewing the entire footage and was suggested only to be used as a fast screening tool. In this paper we propose a method of bleeding detection that uses in its first stage Hue-Saturation-Intensity colour histograms to track a moving background and bleeding colour distributions over time. Such an approach addresses the problem caused by drastic changes in blood colour distribution that occur when it is altered by gastrointestinal fluids and allow detection of other red lesions, which although are usually "less red" than fresh bleeding, they can still be detected when the difference between their colour distributions and the background is large enough. In the second stage of our method, we analyse all candidate blood frames, by extracting colour (HSI) and texture (LBP) features from the suspicious image regions (obtained in the first stage) and their neighbourhoods and classifying them using Support Vector Classifier into Bleeding, Lesion and Normal classes. We show that our algorithm compares favourably with the SBI on the test set of 84 full length videos.

We are developing small animal imaging techniques to characterize the kinetics of lipid accumulation/reduction of fat depots in response to genetic/dietary factors associated with obesity and metabolic syndromes. Recently, we developed an MR ratio imaging technique that approximately yields lipid/{lipid + water}. In this work, we develop a statistical model for the ratio distribution that explicitly includes a partial volume (PV) fraction of fat and a mixture of a Rician and multiple Gaussians. Monte Carlo hypothesis testing showed that our model was valid over a wide range of coefficient of variation of the denominator distribution (c.v.: 0-0:20) and correlation coefficient among the numerator and denominator (&rgr; 0-0.95), which cover the typical values that we found in MRI data sets (c.v.: 0:027-0:063, &rgr;: 0:50-0:75). Then a maximum a posteriori (MAP) estimate for the fat percentage per voxel is proposed. Using a digital phantom with many PV voxels, we found that ratio values were not linearly related to PV fat content and that our method accurately described the histogram. In addition, the new method estimated the ground truth within +1.6% vs. +43% for an approach using an uncorrected ratio image, when we simply threshold the ratio image. On the six genetically obese rat data sets, the MAP estimate gave total fat volumes of 279 ± 45mL, values ≈ 21% smaller than those from the uncorrected ratio images, principally due to the non-linear PV effect. We conclude that our algorithm can increase the accuracy of fat volume quantification even in regions having many PV voxels, e.g. ectopic fat depots.

Osteoarthritis (OA) is a degenerative joint disease characterized by degradation of the articular cartilage, and is a major cause of disability. At present, there is no cure for OA and currently available treatments are directed towards relief of symptoms. Recently it was shown that cartilage homogeneity visualized by MRI and representing the biochemical changes undergoing in the cartilage is a potential marker for early detection of knee OA. In this paper based on homogeneity we present an automatic technique, embedded in a variational framework, for localization of a region of interest in the knee cartilage that best indicates where the pathology of the disease is dominant. The technique is evaluated on 283 knee MR scans. We show that OA affects certain areas of the cartilage more distinctly, and these are more towards the peripheral region of the cartilage. We propose that this region in the cartilage corresponds anatomically to the area covered by the meniscus in healthy subjects. This finding may provide valuable clues in the pathology and the etiology of OA and thereby may improve treatment efficacy. Moreover our method is generic and may be applied to other organs as well.

Image blur introduced by patient motion is one of the most frequently cited reasons for image rejection in radiographic diagnostic imaging. The goal of the present work is to provide an automated method for the detection of anatomical motion blur in digital radiographic images to help improve image quality and facilitate workflow in the radiology department. To achieve this goal, the method first reorients the image to a predetermined hanging protocol. Then it locates the primary anatomy in the radiograph and extracts the most indicative region for motion blur, i.e., the region of interest (ROI). The third step computes a set of motion-sensitive features from the extracted ROI. Finally, the extracted features are evaluated by using a classifier that has been trained to detect motion blur. Preliminary experiments show promising results with 86% detection sensitivity, 72% specificity, and an overall accuracy of 76%.

We introduce a new hybrid physics-based approach for elastic image registration using approximating splines. As underlying deformation model we employ Gaussian elastic body splines (GEBS), which are an analytic solution of the Navier equation under Gaussian forces and are represented by matrix-valued basis functions. Our approach is formulated as an energy-minimizing functional that incorporates both landmark and intensity information as well as a regularization based on GEBS. We also include landmark localization uncertainties represented by weight matrices. Since the approach is based on a physical deformation model, cross-effects in elastic deformations can be handled. We demonstrate the applicability of our scheme based on MR images of the human brain. It turns out that the new scheme is superior to a pure landmark-based as well as a pure intensity-based scheme.

In this paper, we present the latest results of the development of a novel non-rigid image registration method (NiRuDeGG) using a well-established mathematical framework known as the deformation based grid generation. The deformation based grid generation method is able to generate a grid with desired grid density distribution which is free from grid folding. This is achieved by devising a positive monitor function describing the anticipated grid density in the computational domain. Based on it, we have successfully developed a new non-rigid image registration method, which has many advantages. Firstly, the functional to be optimized consists of only one term, a similarity measure. Thus, no regularization functional is required in this method. In particular, there is no weight to balance the regularization functional and the similarity functional as commonly required in many non-rigid image registration methods. Nevertheless, the regularity (no mesh folding) of the resultant deformation is theoretically guaranteed by controlling the Jacobian determinant of the transformation. Secondly, since no regularization term is introduced in the functional to be optimized, the resultant deformation field is highly flexible that large deformation frequently experienced in inter-patient or image-atlas registration tasks can be accurately estimated. Detailed description of the deformation based grid generation, a least square finite element (LSFEM) solver for the underlying div-curl system, and a fast div-curl solver approximating the LSFEM solution using inverse filtering, along with several 2D and 3D experimental results are presented.

We present a novel framework for describing intensity-based multi-modal similarity measures. Our framework is based around a concept of internal, or self, similarity. Firstly the locations of multiple regions or patches which are "similar" to each other are identified within a single image. The term "similar" is used here to represent a generic intra-modal similarity measure. Then if we examine a second image in the same locations, and this image is registered to the first image, we should find that the patches in these locations are also "similar", though the actual features in the patches when compared between the images could be very different. We propose that a measure based on this principle could be used as an inter-modal similarity measure because, as the two images become increasingly misregistered then the patches within the second image should become increasingly dissimilar. Therefore, our framework results in an inter-modal similarity measure by using two intra-modal similarity measures applied separately within each image. In this paper we describe how popular multi-modal similarity measures such as mutual information can be described within this framework. In addition the framework has the potential to allow the formation of novel similarity measures which can register using regional information, rather than individual pixel/voxel intensities. An example similarity measure is produced and its ability to guide a registration algorithm is investigated. Registration experiments are carried out using three datasets. The pairs of images to be registered were specifically chosen as they were expected to challenge (i.e. result in misregistrations) standard intensity-based measures, such as mutual information. The images include synthetic data, cadaver data and clinical data and cover a range of modalities. Our experiments show that our proposed measure is able to achieve accurate registrations where standard intensity-based measures, such as mutual information, fail.

In this paper a method to remove the divergence from a vector field is presented. When applied to a displacement field, this will remove all local compression and expansion. The method can be used as a post-processing step for (unconstrained) registered images, when volume changes in the deformation field are undesired. The method involves solving Poisson's equation for a large system. Algorithms to solve such systems include Fourier analysis and Cyclic Reduction. These solvers are vastly applied in the field of fluid dynamics, to compensate for numerical errors in calculated velocity fields. The application to medical image registration as described in this paper, has to our knowledge not been done before. To show the effect of the method, it is applied to the registration of both synthetic data and dynamic MR series of the liver. The results show that the divergence in the displacement field can be reduced by a factor of 10-1000 and that the accuracy of the registration increases.

The ability of computed tomography (CT) and magnetic resonance (MR) imaging to visualize and discriminate between normal and diseased tissues is improved with contrast agents, which are designed to differentially accumulate in tissues and modify their inherent imaging signal. Conventional contrast agents are limited to a single modality and require fast acquisitions due to rapid clearance following injection. Encapsulation of iohexol and gadoteridol within a nano-engineered liposome has been achieved and can increase their in vivo half-life to several days. We hypothesize that the persistence of this contrast agent in vivo, and the simultaneous co-localized contrast enhancement across modalities will improve longitudinal image registration. This work investigates the in vivo registration performance of the dual-modality contrast agent under realistic conditions. Previous characterizations of single-modality contrast agents were limited to qualitative inspections of signal intensity enhancement. We present quantitative, information theoretic methods for assessing image registration performance. The effect of increased localized contrast upon the mutual information of the MR and CT image sets was shown to increase post-injection. Images registered post- injection had a decreased registration error compared with pre-contrast images. Performance was maintained over extended time frames, contrast agent concentrations, and with decreased field-of-view. This characterization allows optimization of the contrast agent against desired performance for a given imaging task. The ability to perform robust longitudinal image registration is essential for pre-clinical investigations of tumor development, monitoring of therapy response, and therapy guidance over multiple fractions where registration of online cone-beam CT to planning CT and MR is necessary.

We present a new statistical deformation model suited for parameterized grids with different resolutions. Our method models the covariances between multiple grid levels explicitly, and allows for very efficient fitting of the model to data on multiple scales. The model is validated on a data set consisting of 62 annotated MR images of Corpus Callosum. One fifth of the data set was used as a training set, which was non-rigidly registered to each other without a shape prior. From the non-rigidly registered training set a shape prior was constructed by performing principal component analysis on each grid level and using the results to construct a conditional shape model, conditioning the finer parameters with the coarser grid levels. The remaining shapes were registered with the constructed shape prior. The dice measures for the registration without prior and the registration with a prior were 0.875 ± 0.042 and 0.8615 ± 0.051, respectively.

Atherosclerosis at the carotid bifurcation resulting in cerebral emboli is a major cause of ischemic stroke. Most strokes associated with carotid atherosclerosis can be prevented by lifestyle/dietary changes and pharmacological treatments if identified early by monitoring carotid plaque changes. Plaque composition information from magnetic resonance (MR) carotid images and dynamic characteristics information from 3D ultrasound (US) are necessary for developing and validating US imaging tools to identify vulnerable carotid plaques. Combining these images requires nonrigid registration to correct the non-linear miss-alignments caused by relative twisting and bending in the neck due to different head positions during the two image acquisitions sessions. The high degree of freedom and large number of parameters associated with existing nonrigid image registration methods causes several problems including unnatural plaque morphology alteration, computational complexity, and low reliability. Our approach was to model the normal movement of the neck using a "twisting and bending model" with only six parameters for nonrigid registration. We evaluated our registration technique using intra-subject in-vivo 3D US and 3D MR carotid images acquired on the same day. We calculated the Mean Registration Error (MRE) between the segmented vessel surfaces in the target image and the registered image using a distance-based error metric after applying our "twisting bending model" based nonrigid registration algorithm. We achieved an average registration error of 1.33±0.41mm using our nonrigid registration technique. Visual inspection of segmented vessel surfaces also showed a substantial improvement of alignment with our non-rigid registration technique.

Alignment of angiographic preoperative 3D scans to intraoperative 2D projections is an important issue for 3D depth perception and navigation during interventions. Currently, in a setting where only one 2D projection is available, methods employing a rigid transformation model present the state of the art for this problem. In this work, we introduce a method capable of deformably registering 3D vessel structures to a respective single projection of the scene. Our approach addresses the inherent ill-posedness of the problem by incorporating a priori knowledge about the vessel structures into the formulation. We minimize the distance between the 2D points and corresponding projected 3D points together with regularization terms encoding the properties of length preservation of vessel structures and smoothness of deformation. We demonstrate the performance and accuracy of the proposed method by quantitative tests on synthetic examples as well as real angiographic scenes.

In this paper we present a new approach for the registration of cardiac 4D image sequences of different subjects, where we assume that a temporal association between the sequences is given. Moreover, we allow for one (or two) selected pair(s) of associated points in time of both sequences, which we call the bridging points in time, the use of additional information such as the semi-automatic segmentation of the investigated structure. We establish the 3D inter-subject registration for all other pairs of points in time exploiting (1) the inter-subject registration for the bridging pair of points in time, (2) the intra-subject motion calculation in both sequences with respect to the bridging pair, and (3) the concatenation of the obtained transformations. We formulate a cost functional integrating the similarity measures comparing the images of the bridging pair(s) of points in time and of the current pair of points in time, respectively. We evaluated our algorithm on 8 healthy volunteers leading to 28 inter-subject combinations and we analyze the behaviour for different parameter settings weighting differently the involved pairs of points in time. The approach based on the bridging pairs outperforms a direct 3D registration of corresponding points in time, in particular in the right ventricle we gain up to 33% in registration accuracy. Starting with a cost functional taking into account the similarity at the first bridging point in time, the results improve stepwise by integrating, firstly, information from the current pair of points in time and secondly, from a second bridging point in time. Our results also show a steep rise of the importance of regularization on the registration accuracy when registering the current point in time with our procedure (17% gain in accuracy) with respect to a direct registration in the bridging point (less than 1%). However, regularization during intra-sequence registration had only minor effects on the accuracy of our registration procedure.

Blood flow properties in the heart can be examined non invasively by means of Phase Contrast MRI (PC MRI), an imaging technique that provides not only morphology images but also velocity information. We present a novel feature combination for level set segmentation of the heart's cavities in multidirectional 4D PC MRI data. The challenge in performing the segmentation task successfully in this context is first of all the bad image quality, as compared to classical MRI. As generally in heart segmentation, the intra and inter subject variability of the heart has to be coped with as well. The central idea of our approach is to integrate a set of essentially differing sources of information into the segmentation process to make it capable of handling qualitatively bad and highly varying data. To the best of our knowledge our system is the first to concurrently incorporate a flow measure as well as a priori shape knowledge into a level set framework in addition to the commonly used edge and curvature information. The flow measure is derived from PC MRI velocity data. As shape knowledge we use a 3D shape of the respective cavity. We validated our system design by a series of qualitative performance tests. The combined use of shape knowledge and a flow measure increases segmentation quality compared to results obtained by using only one of those features. A first clinical study was performed on two 4D datasets, from which we segmented the left ventricle and atrium. The segmentation results were examined by an expert and judged suitable for use in clinical practice.

This work investigates knowledge driven segmentation of cardiac MR perfusion sequences. We build upon previous work on multi-band AAMs to integrate into the segmentation both spatial priors about myocardial shape as well as temporal priors about characteristic perfusion patterns. Different temporal and spatial features are developed without a strict need for temporal correspondence across the image sequences. We also investigate which combination of spatial and temporal features yields the best segmentation performance. Our evaluation criteria were boundary errors wrt manual segmentations, area overlap, and convergence envelope. From a quantitative evaluation on 19 perfusion studies, we conclude that a combination of the maximum intensity projection feature and gradient orientation map yields the best segmentation performance, with an average point-to-curve error of 0.9-1 pixel wrt manual contours. We also conclude that addition of different temporal features does not necessarily increase performance.

Multi-chamber heart segmentation is a prerequisite for quantification of the cardiac function. In this paper, we propose an automatic heart chamber segmentation system. There are two closely related tasks to develop such a system: heart modeling and automatic model fitting to an unseen volume. The heart is a complicated non-rigid organ with four chambers and several major vessel trunks attached. A flexible and accurate model is necessary to capture the heart chamber shape at an appropriate level of details. In our four-chamber surface mesh model, the following two factors are considered and traded-off: 1) accuracy in anatomy and 2) easiness for both annotation and automatic detection. Important landmarks such as valves and cusp points on the interventricular septum are explicitly represented in our model. These landmarks can be detected reliably to guide the automatic model fitting process. We also propose two mechanisms, the rotation-axis based and parallel-slice based resampling methods, to establish mesh point correspondence, which is necessary to build a statistical shape model to enforce priori shape constraints in the model fitting procedure. Using this model, we develop an efficient and robust approach for automatic heart chamber segmentation in 3D computed tomography (CT) volumes. Our approach is based on recent advances in learning discriminative object models and we exploit a large database of annotated CT volumes. We formulate the segmentation as a two step learning problem: anatomical structure localization and boundary delineation. A novel algorithm, Marginal Space Learning (MSL), is introduced to solve the 9-dimensional similarity transformation search problem for localizing the heart chambers. After determining the pose of the heart chambers, we estimate the 3D shape through learning-based boundary delineation. Extensive experiments demonstrate the efficiency and robustness of the proposed approach, comparing favorably to the state-of-the-art. This is the first study reporting stable results on a large cardiac CT dataset with 323 volumes. In addition, we achieve a speed of less than eight seconds for automatic segmentation of all four chambers.

Segmentation of organs in medical images can be successfully performed with shape-constrained deformable models. A surface mesh is attracted to detected image boundaries by an external energy, while an internal energy keeps the mesh similar to expected shapes. Complex organs like the heart with its four chambers can be automatically segmented using a suitable shape variablility model based on piecewise affine degrees of freedom. In this paper, we extend the approach to also segment highly variable vascular structures. We introduce a dedicated framework to adapt an extended mesh model to freely bending vessels. This is achieved by subdividing each vessel into (short) tube-shaped segments ("tubelets"). These are assigned to individual similarity transformations for local orientation and scaling. Proper adaptation is achieved by progressively adapting distal vessel parts to the image only after proximal neighbor tubelets have already converged. In addition, each newly activated tubelet inherits the local orientation and scale of the preceeding one. To arrive at a joint segmentation of chambers and vasculature, we extended a previous model comprising endocardial surfaces of the four chambers, the left ventricular epicardium, and a pulmonary artery trunk. Newly added are the aorta (ascending and descending plus arch), superior and inferior vena cava, coronary sinus, and four pulmonary veins. These vessels are organized as stacks of triangulated rings. This mesh configuration is most suitable to define tubelet segments. On 36 CT data sets reconstructed at several cardiac phases from 17 patients, segmentation accuracies of 0.61-0.80mm are obtained for the cardiac chambers. For the visible parts of the newly added great vessels, surface accuracies of 0.47-1.17mm are obtained (larger errors are asscociated with faintly contrasted venous structures).

A novel 2D slice based automatic method for model based segmentation of the outer vessel wall of the common carotid artery in CTA data set is introduced. The method utilizes a lumen segmentation and AdaBoost, a fast and robust machine learning algorithm, to initially classify (mark) regions outside and inside the vessel wall using the distance from the lumen and intensity profiles sampled radially from the gravity center of the lumen. A similar method using the distance from the lumen and the image intensity as features is used to classify calcium regions. Subsequently, an ellipse shaped deformable model is fitted to the classification result. The method has achieved smaller detection error than the inter observer variability, and the method is robust against variation of the training data sets.

Knowledge-based vascular segmentation methods typically rely on a pre-built training set of segmented images, which is used to estimate the probability of each voxel to belong to a particular tissue. In 3D Rotational Angiography (3DRA) the same tissue can correspond to different intensity ranges depending on the imaging device, settings and contrast injection protocol. As a result, pre-built training sets do not apply to all images and the best segmentation results are often obtained when the training set is built specifically for each individual image. We present an Image Intensity Standardization (IIS) method designed to ensure a correspondence between specific tissues and intensity ranges common to every image that undergoes the standardization process. The method applies a piecewise linear transformation to the image that aligns the intensity histogram to the histogram taken as reference. The reference histogram has been selected from a high quality image not containing artificial objects such as coils or stents. This is a pre-processing step that allows employing a training set built on a limited number of standardized images for the segmentation of standardized images which were not part of the training set. The effectiveness of the presented IIS technique in combination with a well-validated knowledge-based vasculature segmentation method is quantified on a variety of 3DRA images depicting cerebral arteries and intracranial aneurysms. The proposed IIS method offers a solution to the standardization of tissue classes in routine medical images and effectively improves automation and usability of knowledge-based vascular segmentation algorithms.

Bias field reduction is a common problem in medical imaging. A bias field usually manifests itself as a smooth intensity variation across the image. The resulting image inhomogeneity is a severe problem for posterior image processing and analysis techniques such as registration or segmentation. In this paper, we present a fast debiasing technique based on localized Lloyd-Max quantization. Thereby, the local bias is modelled as a multiplicative field and is assumed to be slowly varying. The method is based on the assumption that the local, undegraded histogram is characterized by a limited number of gray values. The goal is then to find the discrete intensity values such that spreading those values according to the local bias field reproduces the global histogram as good as possible. We show that our method is capable of efficiently reducing (even strong) bias fields in 3D volumes in only a few seconds.

Dermoscopy, also called surface microscopy, is a non-invasive imaging procedure developed for early screening of skin cancer. With recent advance in skin imaging technologies and image processing techniques, there has been increasing interest in computer-aided diagnosis of skin cancer from dermoscopy images. Such diagnosis requires the identification of over one hundred cutaneous morphological features. However, computer procedures designed for extracting and classifying these intricate features can be distracted by the presence of artifacts like hair, ruler markings, and air bubbles. Therefore, reliable artifact removal is an important pre-processing step for improving the performance of computer-aided diagnosis of skin cancer. In this paper, we present a new scheme that automatically detects and removes hairs and ruler markings from dermoscopy images. Moreover, our method also addresses the issue of preserving morphological features during artifact removal. The key components of this method include explicit curvilinear structure detection and modeling, as well as feature guided exemplar-based inpainting. We experiment on a number of dermoscopy images and demonstrate that our method produces superior results compared to existing techniques.

Modern techniques for technical inspection as well as medical diagnostics and therapy in keyhole-surgery scenarios make use of flexible endoscopes. Common to both application fields are very small natural or manmade entry points to the observed scene, as well as the complexity of the hollow itself. These make the use of rigid lens-based endoscopes or tip chip videoscopes impossible. Due to the fact that the fiber-optic image guide of a flexible endoscope introduces a comb structure to the acquired images, many research has been devoted to algorithms for an effective removal of such artifacts. Oftentimes, this research has been motivated by the fact, that the comb structure prevents an application of some well-established methods offered by the computer vision and image processing community. Unfortunately, the performance of the presented approaches are commonly visually evaluated or with respect to proprietary, non-standardized metrics. Thus, the performances of individual algorithms are hard to compare with each other. For this reasons, we propose a performance measure for fiber-optic imaging devices that has been motivated by the physics of optics. In this field, an optical system is frequently described by linear systems theory and the system's quality can be expressed by its transfer function. The determination of this transfer function has been standardized by the ISO for lens based imaging systems and represents a widely accepted measure for the quality of such systems. In this contribution, we present methods that account for fiber-optic imaging systems and thus enable a standardized performance evaluation. Finally, we demonstrate its use by comparing two recent state of the art comb structure removal algorithms, each of them being a representative of a spatial and a frequency domain method, respectively.

This paper presents a framework for mammogram enhancement that is based on a selective enhancement technique. Several enhancement algorithms under this framework are developed, which include weighted mean gray value- and fuzzy cross-over point-based thresholding methods, algorithm fusion, iterative enhancement method, and statistical decision theory-based techniques. Using various abnormal mammograms, the presented algorithms prove to be more robust and yield superior performance when compared with six representative enhancement approaches available in the literature.

Accurate analysis of the hepatic vasculature is of great importance for many medical applications, such as liver surgical planning and diagnosis of tumors and/or vascular diseases. Vessel segmentation is a pivotal step for the morphological and topological analysis of the vascular systems. Physical imaging limitations together with the inherent geometrical complexity of the vessels make the problem challenging. In this paper, we propose a series of methods and techniques that separate and segment the portal vein and the hepatic vein from CT images, and extract the centerlines of both vessel trees. We compare the results obtained with our iterative segmentation-and-reconnection approach with those obtained with a traditional region growing method, and we show that our results are substantially better.

Standard clinical radiological techniques for determining lesion volume changes in interval exams are, as far as we know, quantitatively non-descriptive or approximate at best. We investigate two new registration based methods that help sketch an improved quantitative picture of lesion volume changes in hepatic interval CT exams. The first method, Jacobian Integration, employs a constrained Thin Plate Spline warp to compute the deformation of the lesion of interest over the intervals. The resulting jacobian map of the deformation is integrated to yield the net lesion volume change. The technique is fast, accurate and requires no segmentation, but is sensitive to misregistration. The second scheme uses a Weighted Gray Value Difference image of two registered interval exams to estimate the change in lesion volume. A linear weighting and trimming curve is used to accurately account for the contribution of partial voxels. This technique is insensitive to slight misregistration and useful in analyzing simple lesions with uniform contrast or lesions with insufficient mutual information to allow the computation of an accurate warp. The methods are tested on both synthetic and in vivo liver lesions and results are evaluated against estimates obtained through careful manual segmentation of the lesions. Our findings so far have given us reason to believe that the estimators are reliable. Further experiments on numerous in vivo lesions will probably establish the improved efficacy of these methods in supporting earlier detection of new disease or conversion from stable to progressive disease in comparison to existing clinical estimation techniques.

This paper presents a study of a more accurately propagating deformable contour for outlining the liver in a Computed Tomography image of the abdomen, relying on the idea that a deformable parametric snake will propagate more accurately to the correct edges of an image when applied to textural information of the image as opposed to simple gray level information. The texture information is quantified using a set of Gabor filters and various methods of curve deformation are investigated, including a traditional vector field, gradient vector flow, and an expanding level-set method. Given the relative similarity in gray values of adjacent soft tissues, we found that a deformation algorithm that provides too large a capture range would be easily distracted by nearby values and therefore unsuitable for the particular task of segmenting the liver. Our results demonstrate both a general increase in performance of snake segmentation across the dataset as well as a significant regional improvement in accuracy, particularly in images corresponding with the top of the liver.

An automated segmentation of lung lobes in thoracic CT images is of interest for various diagnostic purposes like the quantification of emphysema or the localization of tumors within the lung. Although the separating lung fissures are visible in modern multi-slice CT-scanners, their contrast in the CT-image often does not separate the lobes completely. This makes it impossible to build a reliable segmentation algorithm without additional information. Our approach uses general anatomical knowledge represented in a geometrical mesh model to construct a robust lobe segmentation, which even gives reasonable estimates of lobe volumes if fissures are not visible at all. The paper describes the generation of the lung model mesh including lobes by an average volume model, its adaptation to individual patient data using a special fissure feature image, and a performance evaluation over a test data set showing an average segmentation accuracy of 1 to 3 mm.

Robust and accurate segmentation of the human airway tree from multi-detector computed-tomography (MDCT) chest scans is vital for many pulmonary-imaging applications. As modern MDCT scanners can detect hundreds of airway tree branches, manual segmentation and semi-automatic segmentation requiring significant user intervention are impractical for producing a full global segmentation. Fully-automated methods, however, may fail to extract small peripheral airways. We propose an automatic algorithm that searches the entire lung volume for airway branches and poses segmentation as a global graph-theoretic optimization problem. The algorithm has shown strong performance on 23 human MDCT chest scans acquired by a variety of scanners and reconstruction kernels. Visual comparisons with adaptive region-growing results and quantitative comparisons with manually-defined trees indicate a high sensitivity to peripheral airways and a low false-positive rate. In addition, we propose a suite of interactive segmentation tools for cleaning and extending critical areas of the automatically segmented result. These interactive tools have potential application for image-based guidance of bronchoscopy to the periphery, where small, terminal branches can be important visual landmarks. Together, the automatic segmentation algorithm and interactive tool suite comprise a robust system for human airway-tree segmentation.

This paper presents a voxel classification based method for segmenting the human airway tree in volumetric computed tomography (CT) images. In contrast to standard methods that use only voxel intensities, our method uses a more complex appearance model based on a set of local image appearance features and Kth nearest neighbor (KNN) classification. The optimal set of features for classification is selected automatically from a large set of features describing the local image structure at several scales. The use of multiple features enables the appearance model to differentiate between airway tree voxels and other voxels of similar intensities in the lung, thus making the segmentation robust to pathologies such as emphysema. The classifier is trained on imperfect segmentations that can easily be obtained using region growing with a manual threshold selection. Experiments show that the proposed method results in a more robust segmentation that can grow into the smaller airway branches without leaking into emphysematous areas, and is able to segment many branches that are not present in the training set.

Respiratory motion is a complicating factor in radiation therapy, tumor ablation, and other treatments of the thorax and upper abdomen. In most cases, the treatment requires a demanding knowledge of the location of the organ under investigation. One approach to reduce the uncertainty of organ motion caused by breathing is to use prior knowledge of the breathing motion. In this work, we extract lung motion fields of seven patients in 4DCT inhale-exhale images using an iterative shape-constrained deformable model approach. Since data was acquired for radiotherapy planning, images of the same patient over different weeks of treatment were available. Although, respiratory motion shows a repetitive character, it is well-known that patient's variability in breathing pattern impedes motion estimation. A detailed motion field analysis is performed in order to investigate the reproducibility of breathing motion over the weeks of treatment. For that purpose, parameters being significant for breathing motion are derived. The analysis of the extracted motion fields provides a basis for a further breathing motion prediction. Patient-specific motion models are derived by averaging the extracted motion fields of each individual patient. The obtained motion models are adapted to each patient in a leave-one-out test in order to simulate motion estimation to unseen data. By using patient-specific mean motion models 60% of the breathing motion can be captured on average.

PET-based tumor volume segmentation techniques are under investigation in recent years due to the increased utilization of FDG-PET imaging in radiation therapy. We have taken the approach of using a Gaussian mixture model (GMM) to model the image intensity distribution of a selected 3D region that completely covers the tumor, called the "analysis region". The modeling is performed with a predetermined number of Gaussian classes and results in a classification of every voxel into one of these classes. The classes are then grouped together to obtain the tumor volume. The only user interaction required is the selection of the "analysis region" and then the algorithm proceeds automatically to initialize the parameters of the different classes and finds the maximum likelihood estimate with expectation maximization. We used 13 clinical and 19 phantom cases to evaluate the precision and accuracy of the segmentation. Reproducibility was within 10% of the average tumor volume estimate and accuracy was ±35% of the true tumor volume and better when compared to two other proposed techniques. The GMM segmentation is extremely user friendly with good precision and accuracy. It has shown great potential to be used in the clinical environment.

The work focuses on a unique medical repository of digital Uterine Cervix images ("Cervigrams") collected by the National Cancer Institute (NCI), National Institute of Health, in longitudinal multi-year studies. NCI together with the National Library of Medicine is developing a unique web-based database of the digitized cervix images to study the evolution of lesions related to cervical cancer. Tools are needed for the automated analysis of the cervigram content to support the cancer research. In recent works, a multi-stage automated system for segmenting and labeling regions of medical and anatomical interest within the cervigrams was developed. The current paper concentrates on incorporating prior-shape information in the cervix region segmentation task. In accordance with the fact that human experts mark the cervix region as circular or elliptical, two shape models (and corresponding methods) are suggested. The shape models are embedded within an active contour framework that relies on image features. Experiments indicate that incorporation of the prior shape information augments previous results.

We present a segmentation algorithm using a statistical deformation model constructed from CT data of adult male pelves coupled to MRI appearance data. The algorithm allows the semi-automatic segmentation of bone for a limited population of MRI data sets. Our application is pelvic bone delineation from pre-operative MRI for image guided pelvic surgery. Specifically, we are developing image guidance for prostatectomies using the daVinci telemanipulator. Hence the use of male pelves only. The algorithm takes advantage of the high contrast of bone in CT data, allowing a robust shape model to be constructed relatively easily. This shape model can then be applied to a population of MRI data sets using a single data set that contains both CT and MRI data. The model is constructed automatically using fluid based non-rigid registration between a set of CT training images, followed by principal component analysis. MRI appearance data is imported using CT and MRI data from the same patient. Registration optimisation is performed using differential evolution. Based on our limited validation to date, the algorithm may outperform segmentation using non-rigid registration between MRI images without the use of shape data. The mean surface registration error achieved was 1.74 mm. The algorithm shows promise for use in segmentation of pelvic bone from MRI, though further refinement and validation is required. We envisage that the algorithm presented could be extended to allow the rapid creation of application specific models in various imaging modalities using a shape model based on CT data.

Due to the high noise and artifacts typically encountered in ultrasound images, segmenting objects from this modality is one of the most challenging tasks in medical image analysis. Model-based approaches like statistical shape models (SSMs) incorporate prior knowledge that supports object detection in case of incomplete evidence from the image data. How well the model adapts to an unseen image is primarily determined by the suitability of the used appearance model, which evaluates the goodness of fit during model evolution. In this paper, we compare two gradient profile models with a region-based approach featuring local histograms to detect the prostate in 3D transrectal ultrasound (TRUS) images. All models are used within an SSM segmentation framework with optimal surface detection for outlier removal. Evaluation was performed using cross-validation on 35 datasets. While the histogram model failed in 10 cases, both gradient models had only 2 failures and reached an average surface distance of 1.16 ± 0.38 mm in comparison with interactively generated reference contours.

Pulmonary vascular tree segmentation has numerous applications in medical imaging and computer-aided diagnosis (CAD), including detection and visualization of pulmonary emboli (PE), improved lung nodule detection, and quantitative vessel analysis. We present a novel approach to pulmonary vessel segmentation based on a fuzzy segmentation concept, combining the strengths of both threshold and seed point based methods. The lungs of the original image are first segmented and a threshold-based approach identifies core vessel components with a high specificity. These components are then used to automatically identify reliable seed points for a fuzzy seed point based segmentation method, namely fuzzy connectedness. The output of the method consists of the probability of each voxel belonging to the vascular tree. Hence, our method provides the possibility to adjust the sensitivity/specificity of the segmentation result a posteriori according to application-specific requirements, through definition of a minimum vessel-probability required to classify a voxel as belonging to the vascular tree. The method has been evaluated on contrast-enhanced thoracic CT scans from clinical PE cases and demonstrates overall promising results. For quantitative validation we compare the segmentation results to randomly selected, semi-automatically segmented sub-volumes and present the resulting receiver operating characteristic (ROC) curves. Although we focus on contrast enhanced chest CT data, the method can be generalized to other regions of the body as well as to different imaging modalities.

The latest generation of spectral optical coherence tomography (OCT) scanners is able to image 3D cross-sectional volumes of the retina at a high resolution and high speed. These scans offer a detailed view of the structure of the retina. Automated segmentation of the vessels in these volumes may lead to more objective diagnosis of retinal vascular disease including hypertensive retinopathy, retinopathy of prematurity. Additionally, vessel segmentation can allow color fundus images to be registered to these 3D volumes, possibly leading to a better understanding of the structure and localization of retinal structures and lesions. In this paper we present a method for automatically segmenting the vessels in a 3D OCT volume. First, the retina is automatically segmented into multiple layers, using simultaneous segmentation of their boundary surfaces in 3D. Next, a 2D projection of the vessels is produced by only using information from certain segmented layers. Finally, a supervised, pixel classification based vessel segmentation approach is applied to the projection image. We compared the influence of two methods for the projection on the performance of the vessel segmentation on 10 optic nerve head centered 3D OCT scans. The method was trained on 5 independent scans. Using ROC analysis, our proposed vessel segmentation system obtains an area under the curve of 0.970 when compared with the segmentation of a human observer.

With many tumor entities, quantitative assessment of lymph node growth over time is important to make therapy choices or to evaluate new therapies. The clinical standard is to document diameters on transversal slices, which is not the best measure for a volume. We present a new algorithm to segment (metastatic) lymph nodes and evaluate the algorithm with 29 lymph nodes in clinical CT images. The algorithm is based on a deformable surface search, which uses statistical shape models to restrict free deformation. To model lymph nodes, we construct an ellipsoid shape model, which strives for a surface with strong gradients and user-defined gray values. The algorithm is integrated into an application, which also allows interactive correction of the segmentation results. The evaluation shows that the algorithm gives good results in the majority of cases and is comparable to time-consuming manual segmentation. The median volume error was 10.1% of the reference volume before and 6.1% after manual correction. Integrated into an application, it is possible to perform lymph node volumetry for a whole patient within the 10 to 15 minutes time limit imposed by clinical routine.

Cell nuclei segmentation is a key issue in automatic cell image analysis for nuclear malignancy. However, due to the complexity of microscopic images, it is usually not easy to obtain satisfied segmentation results, especially on the separation of touching or overlapping nuclei. We propose a method to separate overlapping nuclei whose shapes are similar to ellipses, even if they are tightly clustered and no edge is present where they touch. As a class-specific approach, it introduces a statistical shape model as an extra constraint within the energy functional that measures the homogeneity of regional intensity. The desired contours of each nucleus can be obtained by minimizing this energy functional. The proposed algorithm has been tested on human cervical nuclei images. Experiment results show that our method can separate touching or overlapping ellipse-like nuclei from each other accurately, and the tests on noisy and textured nuclei images also demonstrate its robustness. The resulting segmentation contours are ellipses in different sizes and directions, therefore the shapes of the nuclei have been preserved to a certain degree. The algorithm can be naturally extended to color images, and also has the potential to deal with the separation for overlapping nuclei of other shapes.

Body part recognition based on CT slice images is very important for many applications in PACS and CAD systems. In this paper, we propose a novel approach that can recognize which body part a slice image belongs to robustly. We focus on how to effectively express and use the unique statistical information of the correlation between the CT value and the position information of each body part. We apply the machine learning method AdaBoost to express and use this statistical information. Our approach consists of a training process and a recognition process. In the training process, we first define the whole body using a set of specific classes to ensure that training images in the same class have a high similarity, and prepare a training image set (positive samples and negative samples) for each class. Second, the training images are normalized to a fixed size and rotation in each class. Third, features are calculated for each normalized training image. Finally, AdaBoosted histogram classifiers are trained. After the training process, each class has its own classifiers. In the recognition process, given a series of CT images, the scores of all classes for each slice image are calculated based on the classifiers obtained in the training process. Then, based on the scores of each slice and a simple model of body part sequence continuity, we use dynamic programming (DP) to eliminate false recognition results. Experimental results on 440 unknown series including lesions show that our approach has high a recognition rate.

Biometric technique has been implemented instead of conventional identification methods such as password in computer, automatic teller machine (ATM), and entrance and exit management system. We propose a personal identification (PI) system using color retinal fundus images which are unique to each individual. The proposed procedure for identification is based on comparison of an input fundus image with reference fundus images in the database. In the first step, registration between the input image and the reference image is performed. The step includes translational and rotational movement. The PI is based on the measure of similarity between blood vessel images generated from the input and reference images. The similarity measure is defined as the cross-correlation coefficient calculated from the pixel values. When the similarity is greater than a predetermined threshold, the input image is identified. This means both the input and the reference images are associated to the same person. Four hundred sixty-two fundus images including forty-one same-person's image pairs were used for the estimation of the proposed technique. The false rejection rate and the false acceptance rate were 9.9×10^-5% and 4.3×10^-5%, respectively. The results indicate that the proposed method has a higher performance than other biometrics except for DNA. To be used for practical application in the public, the device which can take retinal fundus images easily is needed. The proposed method is applied to not only the PI but also the system which warns about misfiling of fundus images in medical facilities.

Atlas-based segmentation has proven effective in multiple applications. Usually, several reference images are combined to create a representative average atlas image. Alternatively, a number of independent atlas images can be used, from which multiple segmentations of the image of interest are derived and later combined. One of the major drawbacks of this approach is its large computational burden caused by the high number of required registrations. To address this problem, we introduce One Registration, Multiple Segmentations (ORMS), a procedure to obtain multiple segmentations with a single online registration. This can be achieved by pre-computing intermediate transformations from the initial atlas images to an average image. We show that, compared to the usual approach, our method reduces time considerably with little or no loss in accuracy. On the other hand, optimum combination of these segmentations remains an unresolved problem. Different approaches have been adopted, but they are all far from the upper bound of any combination strategy. This is given by the Combination Oracle, which classifies a voxel correctly if any individual segmentation coincides with the ground truth. We present here a novel combination approach, based on weighting the different segmentations according to the mutual information between the test image and the atlas image after registration. We compare this method with other existing combination strategies using microscopic MR images of mouse brains, achieving statistically significant improvement in segmentation accuracy.

The content of medical images can often be described as a composition of relevant objects with distinct relationships. Each single object can then be represented as a graph node, and local features of the objects are associated as node attributes, e.g. the centroid coordinates. The relations between these objects are represented as graph edges with annotated relational features, e.g. their relative size. Nodes and edges build an attributed relational graph (ARG). For a given setting, e.g. a hand radiograph, a generalization of the relevant objects, e.g. individual bone segments, can be obtained by the statistical distributions of all attributes computed from training images. These yield a structural prototype graph consisting of one attributed node per relevant object and of their relations represented as attributed edges. In contrast to the ARG, the mean and standard deviation of each local or relational feature are used to annotate the prototype nodes or edges, respectively. The prototype graph can then be used to identify the generalized objects in new images. As new image content is represented by hierarchical attributed region adjacency graphs (HARAGs) which are obtained by region-growing, the task of object or scene identification corresponds to the problem of inexact sub-graph matching between a small prototype and the current HARAG. For this purpose, five approaches are evaluated in an example application of bone-identification in 96 radiographs: Nested Earth Mover's Distance, Graph Edit Distance, a Hopfield Neural Network, Pott's Mean Field Annealing and Similarity Flooding. The discriminative power of 34 local and 12 relational features is judged for each object by sequential forward selection. The structural prototypes improve recall by up to 17% in comparison to the approach without relational information.

Texture-based models are intensively used in medical image processing to quantify the homogeneity and consistency of soft tissues across different patients. Several research studies have shown that the co-occurrence texture model and its Haralick descriptors can be successfully applied to capture the statistical properties of the soft tissues' patterns. Given that the calculation of the co-occurrence texture model is a computationally-intensive task, in this paper we investigate the usefulness of using all possible angles and all displacements for capturing the texture properties of an organ of interest, specifically, the liver. Based on the Analysis of Variance (ANOVA) technique and multiple pair-wise comparisons, we found that using only the "near" and "far" displacements is enough to capture the spatial properties of the texture for the liver.

In this paper, we propose a new method for automated detection and segmentation of different tissue types in digitized uterine cervix images using mean-shift clustering and support vector machines (SVM) classification on cluster features. We specifically target the segmentation of precancerous lesions in a NCI/NLM archive of 60,000 cervigrams. Due to large variations in image appearance in the archive, color and texture features of a tissue type in one image often overlap with that of a different tissue type in another image. This makes reliable tissue segmentation in a large number of images a very challenging problem. In this paper, we propose the use of powerful machine learning techniques such as Support Vector Machines (SVM) to learn, from a database with ground truth annotations, critical visual signs that correlate with important tissue types and to use the learned classifier for tissue segmentation in unseen images. In our experiments, SVM performs better than un-supervised methods such as Gaussian Mixture clustering, but it does not scale very well to large training sets and does not always guarantee improved performance given more training data. To address this problem, we combine SVM and clustering so that the features we extracted for classification are features of clusters returned by the mean-shift clustering algorithm. Compared to classification using individual pixel features, classification by cluster features greatly reduces the dimensionality of the problem, thus it is more efficient while producing results with comparable accuracy.

We propose a novel image reconstruction method for dual-energy subtraction radiography. When one of the dual-energy images is obtained at a low dose, a bone image generated with a dual-energy subtraction technique is degraded due to noise, especially high frequency noise. Our method restores the degraded bone image using a regression filter trained by support vector regression. The regression filter is trained based on the input of degraded bone images against an output of corresponding noiseless bone images. Due to strong correlation between the high frequency and low frequency signals of bone, the high frequency signal can be accurately generated based on the observed low frequency signal. However, learning such correlation directly is generally difficult. Therefore our technique first generates a "2-class bone model" that explicitly expresses a bone structure that should be restored. Then while utilizing this model, regression filtering is applied. The accuracy of regression learning is largely improved with this approach. Verification tests show that our method works well: a soft-tissue image obtained by subtracting a restored bone image from a standard radiograph reveals that the rib structure has been thoroughly removed and that the sharpness of the soft-tissue signal is improved in general and among the fine vessels. In conclusion, the proposed method can provide superior dose reduction as well as a better reflection of the anatomical structures in an image. With these advantages, the proposed method can offer high clinical value for the detection of lung lesions.

Recently, several kinds of post-processing image filters which reduce the noise of computed tomography (CT) images have been proposed. However, these image filters are mostly for adults. Because these are not very effective in small (< 20 cm) display fields of view (FOV), we cannot use them for pediatric body images (e.g., premature babies and infant children). We have developed a new noise reduction filter algorithm for pediatric body CT images. This algorithm is based on a 3D post-processing in which the output pixel values are calculated by nonlinear interpolation in z-directions on original volumetric-data-sets. This algorithm does not need the in-plane (axial plane) processing, so the spatial resolution does not change. From the phantom studies, our algorithm could reduce SD up to 40% without affecting the spatial resolution of x-y plane and z-axis, and improved the CNR up to 30%. This newly developed filter algorithm will be useful for the diagnosis and radiation dose reduction of the pediatric body CT images.

Motivated by the goal of improving detection of micro-vessels with low contrast, a new technique based on directional field is presented for enhancing vessels in retinal images. This technique consists of two steps: the estimation of directional field and the enhancement. We will make enhancement along the vascular direction and normalize the brightness of image in a single step. After estimating the directional field, the mean and variance values in a local neighborhood were calculated pixel-by-pixel, accordingly normalize and enhance the local neighborhood. In order to eliminate the artificial boundary between two adjacent areas, an anisotropic Gaussian kernel was introduced to weight the enhancement. The proposed method can obviously increase the contrast of retinal vessels. 20 retinal images were tested in our experiments, and the results demonstrate an effective vessel enhancement algorithm.

This paper describes a fast multi-scale vessel enhancement filter in 3D medical images. For efficient review of the vascular information, clinicians need rendering the 3D vascular information as a 2D image. Generally, the maximum intensity projection (MIP) is a useful and widely used technique for producing a 2D image from the 3D vascular data. However, the MIP algorithm reduces the conspicuousness for small and faint vessels owing to the overlap of non-vascular structures. To overcome this invisibility, researchers have examined the multi-scale vessel enhancement filter based on a combination of the eigenvalues of the 3D Hessian matrix. This multi-scale vessel enhancement filter produces higher contrast. However, it is time-consuming and requires high cost computation due to large volume of data and complex 3D convolution. For fast vessel enhancement, we propose a novel multi-scale vessel enhancement filter using 3D integral images and 3D approximated Gaussian kernel. This approximated kernel looks like cube but it is not exact cube. Each layer of kernel is approximated 2D Gaussian second order derivative by dividing it into three rectangular regions whose sum is integer. 3D approximated kernel is a pile of these 2D box kernels which are normalized by Frobenius norm. Its size fits to vessel width in order to achieve better visualization of the small vessel. Proposed method is approximately five times faster and produces comparable results with previous multi-scale vessel enhancement filter.

Image quality from ultrasound and optical coherence tomography (OCT) is degraded by speckle patterns, which limit the detection of small features and cause a loss of image contrast. To reduce speckle patterns, a novel adaptive kernel algorithm suitable for linearly scaled and log-compressed OCT and ultrasound images is presented. This algorithm combines region growing with a stick based approach. For each direction from the center of a square window to its border, a stick length is selected depending on a homogeneity criterion. Such a set of sticks forms an individual filter kernel for each image pixel. The current kernel size is observed to detect outliers in speckle. If the kernel size drops below a threshold, an outlier is assumed and the filter output is corrected by using a median filter in a second filtering stage. A new homogeneity model is presented that incorporates two existing models and can be fit to actual image statistics. In addition, methods to compute filter parameters for different speckle correlation lengths and imaging systems are presented. An FPGA real-time implementation is proposed and discussed. Measured and simulated speckle images are processed and results compared to existing FPGA based speckle reduction methods. The proposed filter provides good results for various kinds of medical speckle images. The flexible nature of the proposed kernel guarantees suitability for highly correlated as well as for uncorrelated speckle patterns.

Esophageal ultrasound (EUS) is particularly useful for isolating lymph nodes in the N-staging of esophageal cancer, a disease with very poor overall prognosis. Although EUS is relatively low-cost and real time, and it provides valuable information to the clinician, its usefulness to less trained "users" including opportunities for computer-aided diagnosis is still limited due to the strong presence of spatially correlated interference noise called speckles. To this end, in this paper, we present a technique for enhancing lymph nodes in EUS images by first reducing the spatial correlation of the specular noise and then using a modified structured tensor-based anisotropic filter to complete the speckle reduction process. We report on a measure of the enhancement and also on the extent of automatic processing possible, after the speckle reduction process has taken place. Also, we show the limitations of the enhancement process by extracting relevant lymph node features from the despeckled images. When tested on five representative classes of esophageal lymph nodes, we found the despeckling process to greatly reduce the specularity of the original EUS images, therefore proving very useful for visualization purposes. But it still requires additional work for the complete automation of the lymph node characterizing process.

Digital Subtraction Angiography (DSA) is a well-established powerful modality for the visualization of stenosis and blood vessels in general. This paper presents two novel approaches which address image quality. In the first approach we combine anisotropic diffusion with nonlinear normalization. The second approach consists of an introduction of a regularization strategy followed by a classification procedure to improve the enhancement. The performances of two strategies are evaluated based on a database of 73 subjects using SNR, CNR and Tenengrad's metric. Compared with conventional DSA, Eigen's diffusion embedded nonlinear enhancement strategies can improve image quality 95.25% in terms of SNR. The regularization embedded linear enhancement strategy can also improve SNR 51.46% compared with conventional DSA. Similar results are obtained by CNR and Tenengrad's metric measurements. Our system runs on a PC-based workstation using C++ in Windows environment.

Inpainting, a technique originally used to restore film and photographs, is used to disocclude hair from dermascopic images of skin lesions. The technique is compared to the conventional software DullRazor, which uses linear interpolation to perform disocclusion. Comparison was performed by simulating occluding hair on a dermascopic image, applying DullRazor and inpainting and calculating the error induced. Inpainting is found to perform approximately 33% better than DullRazor's linear interpolation, and is more stable under heavy occlusion. The results are also compared to published results from two other alternatives: auto-regressive (AR) model signal extrapolation and band-limited (BL) signal interpolation.

Although various image denoising methods such as PDE-based algorithms have made remarkable progress in the past years, the trade-off between noise reduction and edge preservation is still an interesting and difficult problem in the field of image processing and analysis. A new image denoising algorithm, using a modified PDE model based on pixel similarity, is proposed to deal with the problem. The pixel similarity measures the similarity between two pixels. Then the neighboring consistency of the center pixel can be calculated. Informally, if a pixel is not consistent enough with its surrounding pixels, it can be considered as a noise, but an extremely strong inconsistency suggests an edge. The pixel similarity is a probability measure, its value is between 0 and 1. According to the neighboring consistency of the pixel, a diffusion control factor can be determined by a simple thresholding rule. The factor is combined into the primary partial differential equation as an adjusting factor for controlling the speed of diffusion for different type of pixels. An evaluation of the proposed algorithm on the simulated brain MRI images was carried out. The initial experimental results showed that the new algorithm can smooth the MRI images better while keeping the edges better and achieve higher peak signal to noise ratio (PSNR) comparing with several existing denoising algorithms.

Pyramid based methods in image processing provide a helpful framework for accelerating the propagation of information over large spatial domains, increasing the efficiency for large scale applications. Combined with an anisotropic diffusion scheme tailored to preserve the boundaries at a given level, an efficient way for enhancing large structures in 3D images is presented. In our approach, the partial differential equation defining the evolution of the intensity in the image is solved in an explicit scheme at multiple resolutions in an ascending-descending cycle. Intensity 'flux' between distant voxels is allowed, while preserving borders relative to the scale. Experiments have been performed both with phantoms and with real data from 3D Transrectal Ultrasound Imaging. The effectiveness of the method to remove speckle noise and to enhance large structures such as the prostate has been demonstrated. For instance, using two scales reduces the computation time by 87% as compared to a single scale. Furthermore, we show that the boundaries of the prostate are mainly preserved, by comparing with manually outlined edges.

Wireless capsule endoscopy (WCE) is a new clinical technology permitting the visualization of the small bowel, the most difficult segment of the digestive tract. The major drawback of this technology is the high amount of time for video diagnosis. In this study, we propose a method for informative frame detection by isolating useless frames that are substantially covered by turbid fluids or their contamination with other materials, e.g., faecal, semi-processed or unabsorbed foods etc. Such materials and fluids present a wide range of colors, from brown to yellow, and/or bubble-like texture patterns. The detection scheme, therefore, consists of two stages: highly contaminated non-bubbled (HCN) frame detection and significantly bubbled (SB) frame detection. Local color moments in the Ohta color space are used to characterize HCN frames, which are isolated by the Support Vector Machine (SVM) classifier in Stage-1. The rest of the frames go to the Stage-2, where Laguerre gauss Circular Harmonic Functions (LG-CHFs) extract the characteristics of the bubble-structures in a multi-resolution framework. An automatic segmentation method is designed to extract the bubbled regions based on local absolute energies of the CHF responses, derived from the grayscale version of the original color image. Final detection of the informative frames is obtained by using threshold operation on the extracted regions. An experiment with 20,558 frames from the three videos shows the excellent average detection accuracy (96.75%) by the proposed method, when compared with the Gabor based- (74.29%) and discrete wavelet based features (62.21%).

Nuclear medicine dynamic studies of kidneys, bladder and stomach are important diagnostic tools. Accurate generation of time-activity curves from regions of interest (ROIs) requires that the patient remains motionless for the duration of the study. This is not always possible since some dynamic studies may last from several minutes to one hour. Several motion correction solutions have been explored. Motion correction using external point sources is inconvenient and not accurate especially when motion results from breathing, organ motion or feeding rather than from body motion alone. Centroid-based motion correction assumes that activity distribution is only inside the single organ (without background) and uniform, but this approach is impractical in most clinical studies. In this paper, we present a novel technique of motion correction that first tracks the organ of interest in a dynamic series then aligns the organ. The implementation algorithm for target tracking-based motion correction consists of image preprocessing, target detection, target positioning, motion estimation and prediction, tracking (new search region generation) and target alignment. The targeted organ is tracked from the first frame to the last one in the dynamic series to generate a moving trajectory of the organ. Motion correction is implemented by aligning the organ ROIs in the image series to the location of the organ in the first image. The proposed method of motion correction has been applied to several dynamic nuclear medicine studies including radionuclide cystography, dynamic renal scintigraphy, diuretic renography and gastric emptying scintigraphy.

We propose a system for tracking of human spermatozoa in phase-contrast microscopy image sequences. One of the main aims of a computer-aided sperm analysis (CASA) system is to automatically assess sperm quality based on spermatozoa motility variables. In our case, the problem of assessing sperm quality is cast as a multi-object tracking problem, where the objects being tracked are the spermatozoa. The system combines a particle filter and Kalman filters for robust motion estimation of the spermatozoa tracks. Further, the combinatorial aspect of assigning observations to labels in the particle filter is formulated as a linear assignment problem solved using the Hungarian algorithm on a rectangular cost matrix, making the algorithm capable of handling missing or spurious observations. The costs are calculated using hidden Markov models that express the plausibility of an observation being the next position in the track history of the particle labels. Observations are extracted using a scale-space blob detector utilizing the fact that the spermatozoa appear as bright blobs in a phase-contrast microscope. The output of the system is the complete motion track of each of the spermatozoa. Based on these tracks, different CASA motility variables can be computed, for example curvilinear velocity or straight-line velocity. The performance of the system is tested on three different phase-contrast image sequences of varying complexity, both by visual inspection of the estimated spermatozoa tracks and by measuring the mean squared error (MSE) between the estimated spermatozoa tracks and manually annotated tracks, showing good agreement.

Difficulty swallowing, or dysphagia, has become a growing problem. Swallowing complications can lead to malnutrition, dehydration, respiratory infection, and even death. The current gold standard for analyzing and diagnosing dysphagia is the videofluoroscopic barium swallow study. In these studies, a fluoroscope is used to image the patient ingesting barium solutions of different volumes and viscosities. The hyoid bone anchors many key muscles involved in swallowing and plays a key role in the process. Abnormal hyoid bone motion during a swallow can indicate swallowing dysfunction. Currently in clinical settings, hyoid bone motion is assessed qualitatively, which can be subject to intra-rater and inter-rater bias. This paper presents a semi-automatic method for tracking the hyoid bone that makes quantitative analysis feasible. The user defines a template of the hyoid on one frame, and this template is tracked across subsequent frames. The matching phase is optimized by predicting the position of the template based on kinematics. An expert speech pathologist marked the position of the hyoid on each frame of ten studies to serve as the gold standard. Results from performing Bland-Altman analysis at a 95% confidence interval showed a bias of 0.0±0.08 pixels in x and -0.08±0.09 pixels in y between the manually-defined gold standard and the proposed method. The average Pearson's correlation between the gold standard and the proposed method was 0.987 in x and 0.980 in y. This paper also presents a method for automatically establishing a patient-centric coordinate system for the interpretation of hyoid motion. This coordinate system corrects for upper body patient motion during the study and identifies superior-inferior and anterior-posterior motion components. These tools make the use of quantitative hyoid motion analysis feasible in clinical and research settings.

Quantification of diffusion tensor imaging (DTI) parameters has become an important role in the neuroimaging, neurosurgical, and neurological community as a method to identify major white matter tracts afflicted by pathology or tracts at risk for a given surgical approach. We introduce a novel framework for a reliable and robust quantification of DTI parameters, which overcomes problems of existing techniques introduced by necessary user inputs. In a first step, a hybrid clustering method is proposed that allows for extracting specific fiber bundles in a robust way. Compared to previous methods, our approach considers only local proximities of fibers and is insensitive to their global geometry. This is very useful in cases where a fiber tracking of the whole brain is not available. Our technique determines the overall number of clusters iteratively using a eigenvalue thresholding technique to detect disjoint clusters of independent fiber bundles. Afterwards, possible finer substructures based on an eigenvalue regression are determined within each bundle. In a second step, a quantification of DTI parameters of the extracted bundle is performed. We propose a method that automatically determines a 3D image where the voxel values encode the minimum distance to a reconstructed fiber. This image allows for calculating a 3D mask where each voxel within the mask corresponds to a voxel that lies in an isosurface around the fibers. The mask is used for an automatic classification between tissue classes (fiber, background, and partial volume) so that the quantification can be performed on one or more of such classes. This can be done per slice or a single DTI parameter can be determined for the whole volume which is covered by the isosurface. Our experimental tests confirm that major white matter fiber tracts may be robustly determined and can be quantified automatically. A great advantage of our framework is its easy integration into existing quantification applications so that uncertainties can be reduced, and higher intrarater- as well as interrater reliabilities can be achieved.

In this paper, we describe a method to evaluate similarities in estimated temporal noise spectra of functional Magnetic Resonance Imaging (fMRI) time series. Accurate noise spectra are needed for reliable activation detection in fMRI. Since these spectra are a-priori unknown, they have to be estimated from the fMRI data. A noise model can be estimated for each voxel separately, but when noise spectra of neighboring voxels are (almost) equal, the power of the activation detection test can be improved by estimating the noise model from a set of neighboring voxels. In this paper, a method is described to evaluate the similarity of noise spectra of neighboring voxels. Noise spectrum similarities are studied in simulation as well as experimental fMRI datasets. The parameters of the model describing the voxel time series are estimated by a Maximum Likelihood (ML) estimator. The similarity of the ML estimated noise processes is assessed by the Model Error (ME), which is based on the Kullback Leibler divergence. Spatial correlations in the fMRI data reduce the ME between the noise spectra of (neighboring) voxels. This undesired effect is quantified by simulation experiments where spatial correlation is introduced. By plotting the ME as a function of the distance between voxels, it is observed that the ME increases as a function of this distance. Additionally, by using the theoretical distribution of the ME, it is observed that neighboring voxels indeed have similar noise spectra and these neighbors can be used to improve the noise model estimate.

This paper proposes a novel and robust approach to the registration (matching) of intra-subject white matter (WM) fiber sets extracted from DT-MRI scans by Tractography. For each fiber, a feature space representation is obtained by appending the sequence of its 3D coordinates. Clustering by non-parametric adaptive mean shift provides a representative fiber for each cluster hereafter termed the fiber-mode (FM). For each FM, the parameters of a multivariate Gaussian are computed from its fiber population, leading to a mixture of Gaussians (MoG) for the whole fiber set. The number of Gaussians used for a fiber set equals the number of FM representing the set. The alignment of two fiber sets is then treated as the alignment between two MoGs, and is solved by maximizing the correlation ratio between them. Initial results are presented for real intrasubject fiber sets and synthetic transformations.

Diffusion weighted MR imaging is a powerful tool that can be employed to study white matter microstructure by examining the 3D displacement profile of water molecules in brain tissue. By applying diffusion-sensitizing gradients along a minimum of 6 directions, second-order tensors (represetnted by 3-by-3 positive definiite matrices) can be computed to model dominant diffusion processes. However, it has been shown that conventional DTI is not sufficient to resolve more complicated white matter configurations, e.g. crossing fiber tracts. More recently, High Angular Resolution Diffusion Imaging (HARDI) seeks to address this issue by employing more than 6 gradient directions. To account for fiber crossing when analyzing HARDI data, several methodologies have been introduced. For example, q-ball imaging was proposed to approximate Orientation Diffusion Function (ODF). Similarly, the PAS method seeks to reslove the angular structure of displacement probability functions using the maximum entropy principle. Alternatively, deconvolution methods extract multiple fiber tracts by computing fiber orientations using a pre-specified single fiber response function. In this study, we introduce Tensor Distribution Function (TDF), a probability function defined on the space of symmetric and positive definite matrices. Using calculus of variations, we solve for the TDF that optimally describes the observed data. Here, fiber crossing is modeled as an ensemble of Gaussian diffusion processes with weights specified by the TDF. Once this optimal TDF is determined, ODF can easily be computed by analytical integration of the resulting displacement probability function. Moreover, principle fiber directions can also be directly derived from the TDF.

Diffusion Tensor Magnetic Resonance Imaging (DTI) is a well known technique that can provide information about the neuronal fiber structure of the brain. However, since DTI requires a large amount of data, a high speed MRI acquisition technique is needed to acquire these data within a reasonable time. Echo Planar Imaging (EPI) is a technique that provides the desired speed. Unfortunately, the advantage of speed is overshadowed by image artifacts, especially at high fields. EPI artifacts originate from susceptibility differences in adjacent tissues and correction techniques are required to obtain reliable images. In this work, the fieldmap method, which tries to measure distortion effects, is optimized by using a non linear least squares estimator for calculating pixel shifts. This method is tested on simulated data and proves to be more robust against noise compared to previously suggested methods. Another advantage of this new method is that other parameters like relaxation and the odd/even phase difference are estimated. This new way of estimating the field map is demonstrated on a hardware phantom, which consists of parallel bundles made of woven strands of Micro Dyneema fibers. Using a modified EPI-sequence, reference data was measured for the calculation of fieldmaps. This allows one to reposition the pixels in order to obtain images with less distortions. The correction is applied to non-diffusion weighted images as well as diffusion weighted images and fiber tracking is performed on this corrected data.

A multi-tensor model with identifiable parameters is developed for diffusion weighted MR images. A new parameterization method guarantees the symmetric positive-definiteness of the diffusion tensor. We set up a Bayesian method for parameter estimation. To investigate properties of the method, Monte Carlo simulated data from three distinct DTI direction schemes have been analyzed. The multi-tensor model with automatic model selection has also been applied to a healthy human brain dataset. Standard tensor-derived maps are obtained when the single-tensor model is fitted to a region of interest with a single dominant fiber direction. High anisotropy diffusion flows and main diffusion directions can be shown clearly in the FA map and diffusion ellipsoid map. For another region containing crossing fiber bundles, we estimate and display the ellipsoid map under the single tensor and double-tensor regimes of the multi-tensor model, suitably thresholding the Bayes factor for model selection.

Many neuroimaging studies are applied to monkeys as pathologies and environmental exposures can be studied in well-controlled settings and environment. In this work, we present a framework for the use of an atlas based, fully automatic segmentation of brain tissues, lobar parcellations, subcortical structures and the regional extraction of Diffusion Tensor Imaging (DTI) properties. We first built a structural atlas from training images by iterative, joint deformable registration into an unbiased average image. On this atlas, probabilistic tissue maps, a lobar parcellation and subcortical structures were determined. This information is applied to each subjects structural image via affine, followed by deformable registration. The affinely transformed atlas is employed for a joint T1 and T2 based tissue classification. The deformed parcellation regions mask the tissue segmentations to define the parcellation for white and gray matter separately. Each subjects structural image is then non-rigidly matched with its DTI image by normalized mutual information, b-spline based registration. The DTI property histograms were then computed using the probabilistic white matter information for each lobar parcellation. We successfully built an average atlas using a developmental training datasets of 18 cases aged 16-34 months. Our framework was successfully applied to over 50 additional subjects in the age range of 9 70 months. The probabilistically weighted FA average in the corpus callosum region showed the largest increase over time in the observed age range. Most cortical regions show modest FA increase, whereas the cerebellums FA values remained stable. The individual methods used in this segmentation framework have been applied before, but their combination is novel, as is their application to macaque MRI data. Furthermore, this is the first study to date looking at the DTI properties of the developing macaque brain.

An interpolation model is a necessary ingredient of intensity-based registration methods. The properties of such a model depend entirely on its basis function, which has been traditionally characterized by features such as its order of approximation and its support. However, as has been recently shown, these features are blind to the amount of registration bias created by the interpolation process alone; an additional requirement that has been named constant-variance interpolation is needed to remove this bias. In this paper, we present a theoretical investigation of the role of the interpolation basis in a registration context. Contrarily to published analyses, ours is deterministic; it nevertheless leads to the same conclusion, which is that constant-variance interpolation is beneficial to image registration. In addition, we propose a novel family of interpolation bases that can have any desired order of approximation while maintaining the constant-variance property. Our family includes every constant-variance basis we know of. It is described by an explicit formula that contains two free functional terms: an arbitrary 1-periodic binary function that takes values from {-1, 1}, and another arbitrary function that must satisfy the partition of unity. These degrees of freedom can be harnessed to build many family members for a given order of approximation and a fixed support. We provide the example of a symmetric basis with two orders of approximation that is supported over [-3/2, 3/2] this support is one unit shorter than a basis of identical order that had been previously

Efficient retrieval of high quality Regions-Of-Interest (ROI) from high resolution medical images is essential for reliable interpretation and accurate diagnosis. Random access to high quality ROI from codestreams is becoming an essential feature in many still image compression applications, particularly in viewing diseased areas from large medical images. This feature is easier to implement in block based codecs because of the inherent spatial independency of the code blocks. This independency implies that the decoding order of the blocks is unimportant as long as the position for each is properly identified. In contrast, wavelet-tree based codecs naturally use some interdependency that exploits the decaying spectrum model of the wavelet coefficients. Thus one must keep track of the decoding order from level to level with such codecs. We have developed an innovative multi-rate image subband coding scheme using "Backward Coding of Wavelet Trees (BCWT)" which is fast, memory efficient, and resolution scalable. It offers far less complexity than many other existing codecs including both, wavelet-tree, and block based algorithms. The ROI feature in BCWT is implemented through a transcoder stage that generates a new BCWT codestream containing only the information associated with the user-defined ROI. This paper presents an efficient technique that locates a particular ROI within the BCWT coded domain, and decodes it back to the spatial domain. This technique allows better access and proper identification of pathologies in high resolution images since only a small fraction of the codestream is required to be transmitted and analyzed.

Medical imaging is essential in the diagnosis of atherosclerosis. In this paper, we propose the semi-automatic matching of two promising and complementary intravascular imaging techniques, Intravascular Ultrasound (IVUS) and Optical Coherence Tomography (OCT), with the ultimate goal of producing hybrid images with increased diagnostic value for assessing arterial health. If no ECG gating has been performed on the IVUS and OCT pullbacks, there is typically an anatomical shuffle (displacement in time and space) in the image sequences due to the catheter motion in the artery during the cardiac cycle, and thus, this is not possible to perform a 3D registration. Therefore, the goal of our work is to detect semi-automatically the corresponding images in both modalities as a preprocessing step for the fusion. Our method is based on the characterization of the lumen shape by a set of Gabor Jets features. We also introduce different correction terms based on the approximate position of the slice in the artery. Then we train different support vector machines based on these features to recognize these correspondences. Experimental results demonstrate the usefulness of our approach, which achieves up to 95% matching accuracy for our data.

Linear elastic model widely applied for simulation of soft tissue deformations in biomedical imaging applications is basically limited to the range of small deformations and rotations. Thus, computation of large deformations and rotations using linear elastic approximation and its derivatives is associated with substantial error. More realistic modeling of mechanical behavior of soft tissue requires handling of different types of nonlinearities. This paper presents a framework for more accurate modeling of deformable structures based on the St. Venant-Kirchhoff law with the nonlinear Green-Lagrange strain tensor and variable material constants, which considers both material and geometric nonlinearities. We derive the governing partial differential equation of nonlinear elasticity, which represents consistent extension of the Lame-Navier PDE of linear elasticity, and describe two alternative numerical schemes for solving this nonlinear PDE via the Newton's and fixed point method, respectively. The results of our comparative studies demonstrate the advantages of nonlinear elastic model for accurate computing of large deformations and rotations in comparison to the linear elastic approximation.

We have compared and validated image registration methods with respect to the clinically relevant use-case of lung CT max-inhale to max-exhale registration. Four fundamentally different algorithms representing main approaches for image registration were compared using clinical images. Each algorithm was assigned to a different person with extensive working knowledge of its usage. Quantitative and qualitative evaluation is performed. Whereas the methods achieve similar results in target registration error, characteristic differences come to show by closer analysis of the displacement fields.

This work describes a non-rigid registration method for open 2D manifold embedded in 3D Euclidian space. The method is based on difference of distance maps and grid based warps interpolated by splines constrained in such a way that the deformation field is diffeomorphic. We then create a dense surface to surface correspondence using angle weighted normals and ray tracing. The implementation using a derivation of the inverse compositional algorithm for optimization of computational speed is described. The results are evaluated as a shape model showing the principal modes of variation.

We propose a novel registration method, which combines well-known vessel detection techniques with aspects of model adaptation. The proposed method is tailored to the requirements of 2D-3D-registration of interventional angiographic X-ray data such as acquired during abdominal procedures. As prerequisite, a vessel centerline is extracted out of a rotational angiography (3DRA) data set to build an individual model of the vascular tree. Following the two steps of local vessel detection and model transformation the centerline model is matched to one dynamic subtraction angiography (DSA) target image. Thereby, the in-plane position and the 3D orientation of the centerline is related to the vessel candidates found in the target image minimizing the residual error in least squares manner. In contrast to feature-based methods, no segmentation of the vessel tree in the 2D target image is required. First experiments with synthetic angiographies and clinical data sets indicate that matching with the proposed model-to-image based registration approach is accurate and robust and is characterized by a large capture range.

We address the problem of intra-subject registration for change detection. The goal is to separate stationary and changing subsets to be able to robustly perform rigid registration on the stationary subsets and thus improve the subsequent change detection. An iterative approach using a hybrid of parametric and non-parametric statistics is presented. The method uses non-parametric clustering and large scale hypothesis testing with estimation of the empirical null hypothesis. The method is successfully applied to 3D surface scans of human ear impressions containing true changes as well as data with synthesized changes. It is shown that the method improves registration and is capable of reducing the difference between registration using different norms.

To align patients in radiation devices in six degrees of freedom (DoF), image-guided approaches perform the task of correction computation for the patient position. Digital radiography (DR) images are compared to projections of a CT series to estimate misalignments. A problem is that digital reconstructed radiographs (DRRs) have to be created from the CT to be registered with the DRs. Depending on the X-ray tube energy, detector sensitivity and body part involved, DRRs and DRs may look very different and often cannot be registered. We present a method that reconstructs multi-spectral DRRs for different X-ray settings, which can be registered to real X-ray images. As short rendering times are crucial, multiple spectra of a DRR are generated in one ray-tracing process. We register our multi-spectral DRR with the DR and add a further DoF to find a best match not only for the translations and in-plane rotation, but also the best fitting spectral planes. The results are used to identify patient misalignments and show that higher reliability can be achieved compared to conventional approaches. Misalignments can be identified even if ineligible X-ray settings have been used. As our approach allows application of lower X-ray energies for DR creation, an additional benefit is the reduction of the delivered dose.

In this paper, we propose three novel and important methods for the registration of histological images for 3D reconstruction. First, possible intensity variations and nonstandardness in images are corrected by an intensity standardization process which maps the image scale into a standard scale where the similar intensities correspond to similar tissues meaning. Second, 2D histological images are mapped into a feature space where continuous variables are used as high confidence image features for accurate registration. Third, we propose an automatic best reference slice selection algorithm that improves reconstruction quality based on both image entropy and mean square error of the registration process. We demonstrate that the choice of reference slice has a significant impact on registration error, standardization, feature space and entropy information. After 2D histological slices are registered through an affine transformation with respect to an automatically chosen reference, the 3D volume is reconstructed by co-registering 2D slices elastically.

Helmholtz's theorem states that, with suitable boundary condition, a vector field is completely determined if both of its divergence and curl are specified everywhere. Based on this, we developed a new parametric non-rigid image registration algorithm. Instead of the displacements of regular control grid points, the curl and divergence at each grid point are employed as the parameters. The closest related work was done by Kybic where the parameters are the Bspline coefficients of the displacement field at each control grid point. However, in Kybic's work, it is very likely to result in grid folding in the final deformation field if the distance between adjacent control grid points (knot spacing) is less than 8. This implies that the high frequency components in the deformation field can not be accurately estimated. Another relevant work is the NiRuDeGG method where by solving a div-curl system, an intermediate vector field is generated and, in turn, a well-regularized deformation field can be obtained. Though the present work does not guarantee the regularity (no mesh folding) of the resulting deformation field, which is also suffered by Kybic's work, it allows for a more efficient optimization scheme over the NiRuDeGG method. Our experimental results showed that the proposed method is less prone to grid folding than Kybic's work and that in many cases, in a multi-resolution fashion; the knot spacing can be reduced down to 1 and thus has the potential to achieve higher registration accuracy. Detailed comparison among the three algorithms is described in the paper.

Prostate cancer is a multifocal disease and lesions are not distributed uniformly within the gland. Several biopsy protocols concerning spatially specific targeting have been reported urology literature. Recently a statistical cancer atlas of the prostate was constructed providing voxelwise probabilities of cancers in the prostate. Additionally an optimized set of biopsy sites was computed with 94 - 96% detection accuracy was reported using only 6-7 needles. Here we discuss the warping of this atlas to prostate segmented side-fire ultrasound images of the patient. A shape model was used to speed up registration. The model was trained from over 38 expert segmented subjects off-line. This training yielded as few as 15-20 degrees of freedom that were optimized to warp the atlas surface to the patient's ultrasound image followed by elastic interpolation of the 3-D atlas. As a result the atlas is completely mapped to the patient's prostate anatomy along with optimal predetermined needle locations for biopsy. These do not preclude the use of additional biopsies if desired. A color overlay of the atlas is also displayed on the ultrasound image showing high cancer zones within the prostate. Finally current biopsy locations are saved in the atlas space and may be used to update the atlas based on the pathology report. In addition to the optimal atlas plan, previous biopsy locations and alternate plans can also be stored in the atlas space and warped to the patient with no additional time overhead.

Non-rigid multi-modal volume registration is computationally intensive due to its high-dimensional parameter space, where common CPU computation times are several minutes. Medical imaging applications using registration, however, demand ever faster implementations for several purposes: matching the data acquisition speed, providing smooth user interaction and steering for quality control, and performing population registration involving multiple datasets. Current GPUs offer an opportunity to boost the registration speed through high computational power at low cost. In our previous work, we have presented a GPU implementation of a non-rigid multi-modal volume registration that was 6 - 8 times faster than a software implementation. In this paper, we extend this work by describing how new features of the DX10-compatible GPUs and additional optimization strategies can be employed to further improve the algorithm performance. We have compared our optimized version with the previous version on the same GPU, and have observed a speedup factor of 3.6. Compared with the software implementation, we achieve a speedup factor of up to 44.

Registration of 3D images acquired from different imaging modalities such as micro-magnetic resonance imaging (µMRI) and micro-computed tomography (µCT) are of interest in a number of medical imaging applications. Most general-purpose multimodality registration algorithms tend to be computationally intensive and do not take advantage of the shape of the imaging volume. Multimodality trabecular bone (TB) images of cylindrical cores, for example, tend to be misaligned along and around the axial direction more than that around other directions. Additionally, TB images acquired by µMRI can differ substantially from those acquired by µCT due to apparent trabecular thickening from magnetic susceptibility boundary effects and non-linear intensity correspondence. However, they share very similar contrast characteristics since the images essentially represent a binary tomographic system. The directional misalignment and the fundamental similarities of the two types of images can be exploited to achieve fast 3D registration. Here we present an intensity cross-correlation based 3D registration algorithm for registering 3D specimen images from cylindrical cores of cadaveric TB acquired by µMRI and µCT in the context of finite-element modeling to assess the bone's mechanical constants. The algorithm achieves the desired registration by first coarsely approximating the three translational and three rotational parameters required to align the µMR images to the µCT scan coordinate frame and fine-tuning the parameters in the neighborhood of the approximate solution. The algorithm described here is suitable for 3D rigid-body image registration applications where through-plane rotations are known to be relatively small. The accuracy of the technique is constrained by the image resolution and in-plane angular increments used.

We propose a method for combining histopathology image with MR image of the prostate by using image correction and nonrigid registration. Our method consists of four steps. First, two or four tissue sections of the prostate in histopathology image are combined to produce a single prostate image by stitching. Second, the intensity of prostate bleeding area on T2-weighted MR image is substituted for that on T1-weighted MR image. Our intensity correction prevents a mistake which a prostate bleeding is considered as a tumor on T2-weighted MR image. Third, rough and fine registration is performed to find the best match for pixel overlap between histopathology and MR images. Then the result of rigid registration is deformed by the TPS warping. Finally, aligned images are visualized by the intensity intermixing. Experimental results show that the prostate tumor lesion can be properly located and clearly visualized within MR images for tissue characterization comparison.

Three-dimensional datasets of complex objects are readily available from the tomographic modalities, and fusion of these data sets leads to new understanding of the data. Automatic alignment of the objects is difficult or time consuming when substantial misalignments are present or point correspondences cannot be established, or the solution space is non-convex. These issues effectively exclude most optimization algorithms used in conventional data alignment. Here, we present the particle swarm optimization (PSO) approach which is not sensitive to initial conditions, local minima or non-convex solution space. Intercommunicating particle swarms are randomly placed in the solution space (representing the parameters of the rigid transformations). Each member of each swarm traverses the solution space, constantly evaluating the objective function at its own position and communicating with other members of the swarm about theirs. In addition, the swarms communicate between themselves. Through this information sharing between swarm members and the swarms, the space is searched completely and efficiently, and as a result all swarms converge near the globally optimal rigid transformation. To evaluate the technique, high-resolution micro-CT data sets of single mouse heads were acquired with large initial misalignments. Using two communicating particle swarms in the same solution space, six distinct mouse head objects were aligned finding the approximate global minima in about 25 iterations or 140 seconds on a standard PC independent of initial conditions. Faster speeds (better accuracy) can be obtained by relaxing (restricting) the convergence criteria. These results indicate that the particle swarm approach may be a valuable tool for stand-alone or hybrid alignments.

Tracer kinetic modeling with dynamic contrast enhanced MRI (DCE-MRI) and the quantification of the kinetic parameters are active fields of research which have the potential to improve the measurement of renal function. However, the strong coronal motion of the kidney in the time series inhibits an accurate assessment of the kinetic parameters. Automatic motion correction is challenging due to the large movement of the kidney and the strong intensity changes caused by the injected bolus. In this work, we improve the quantification results by a template matching motion correction method using a gradient-based similarity measure. Thus, a tedious manual motion correction is replaced by an automatic procedure. The only remaining user interaction is reduced to a selection of a reference slice and a coarse manual segmentation of the kidney in this slice. These steps do not present an overhead to the interaction needed for the assessment of the kinetic parameters. In order to achieve reliable and fast results, we constrain the degrees of freedom for the correction method as far as possible. Furthermore, we compare our method to deformable registration using the same similarity measure. In all our tests, the presented template matching correction was superior to the deformable approach in terms of reliability, leading to more accurate parameter quantification. The evaluation on 10 patient data series with 180-230 images each demonstrate that the quantitative analysis by a two-compartment model can be improved by our method.

The mid-sagittal plane (MSP) is a commonly used anatomic landmark for standardized MR brain acquisition. In addition to the requirement of accurate detection of the MSP geometry, it is also imperative from clinical point of view to consistently prescribe scan planning for evaluation of pathology process in follow-up studies. In this work, an adaptive technique of scan planning has been developed to enforce the consistency among scans acquired at different time points from the same patient by maximizing image similarity in the proximity of MSP. The geometry parameters of the MSP of current study are optimized by simplex algorithm to achieve better similarity to the reference study. Meanwhile different similarity measures are studied and evaluated within the region of the interest of each MSP. The method is successfully tested on self-reference consistency study by manually setting the reference sagittal image. It is also tested with clinical follow-up studies of MR images acquired from 30 patients. By visual inspection, the adaptive consistency method improves the similarity to the reference images in 22 follow-up studies evidently, while the similarity to the reference images in 7 studies improves slightly. This result demonstrates the efficacy of our method on consistent detection of mid-sagittal planes for follow-up MR brain study.

Our work presents a rapid and robust process that can analytically evaluate and correct patient setup error for head and neck radiotherapy by comparing orthogonal megavoltage portal images with digitally reconstructed radiographs. For robust data Photoshop is used to interactively segment images and registering reference contours to the transformed PI. MatLab is used for matrix computations and image analysis. The closest point distance for each PI point to a DRR point forms a set of homologous points. The translation that aligns the PI to the DRR is equal to the difference in centers of mass. The original PI points are transformed and the process repeated with an Iterative Closest Point algorithm until the transformation change becomes negligible. Using a 3.00 GHz processor the calculation of the 2500x1750 CPD matrix takes about 150 sec per iteration. Standard down sampling to about 1000 DRR and 250 PI points significantly reduces that time. We introduce a local neighborhood matrix consisting of a small subset of the DRR points in the vicinity of each PI point to further reduce the CPD matrix size. Our results demonstrate the effects of down sampling on accuracy. For validation, analytical detailed results are displayed as a histogram.

Fundus camera imaging can be used to examine the retina to detect disorders. Similar to looking through a small keyhole into a large room, imaging the fundus with an ophthalmologic camera allows only a limited view at a time. Thus, the generation of a retinal montage using multiple images has the potential to increase diagnostic accuracy by providing larger field of view. A method of mosaicing multiple retinal images using the radial distortion correction (RADIC) model is proposed in this paper. Our method determines the inter-image connectivity by detecting feature correspondences. The connectivity information is converted to a tree structure that describes the spatial relationships between the reference and target images for pairwise registration. The montage is generated by cascading pairwise registration scheme starting from the anchor image downward through the connectivity tree hierarchy. The RADIC model corrects the radial distortion that is due to the spherical-to-planar projection during retinal imaging. Therefore, after radial distortion correction, individual images can be properly mapped onto a montage space by a linear geometric transformation, e.g. affine transform. Compared to the most existing montaging methods, our method is unique in that only a single registration per image is required because of the distortion correction property of RADIC model. As a final step, distance-weighted intensity blending is employed to correct the inter-image differences in illumination encountered when forming the montage. Visual inspection of the experimental results using three mosaicing cases shows our method can produce satisfactory montages.

To accommodate the inter- and intra-fractional motion of internal organs in prostate cancer treatment, a large margin (5mm-25mm) has often to be considered during radiation therapy planning. Normally, the inter-fractional motion is more substantial than the intra-fractional counterpart. Therefore, the study of inter-fractional motion pattern is of special interest for adaptive radiation therapy. Existing methods on organ motion analysis mainly focus on the deviation of an organ's shape from its mean shape. The deviation information is helpful in choosing a statistically proper margin, but is of limited use for plan adaptation. In this paper, we propose a new deformation analysis method that can be directly used for plan adaptation. First, deformation estimation is accomplished by a fast deformable registration method, which utilizes a contour based multi-grid strategy to register treatment cone-beam CT (CBCT) images with planning CT images. Second, dominant deformation modes are extracted by a novel deformation analysis approach. To be specific, a cooperative principal component analysis (PCA) method is developed to analyze the deformation field in a coarse-to-fine strategy. The deformation modes are initialized by applying PCA on the organs as a whole and refined by analyzing the individual organs subsequently. The experimental results show that the organ motion can be well characterized by a few dominant deformation modes. Based on the dominant modes, a corresponding set of dominant modal plans could be generated for further optimization. Ultimately, an adaptive plan for each treatment can be obtained on-line while the margin can be effectively reduced to minimize the unnecessary radiation dosage.

Spatiotemporal image data sets, like 4D CT or dynamic MRI, open up the possibility to estimate respiratory induced tumor and organ motion and to generate four-dimensional models that describe the temporal change in position and shape of structures of interest. However, two main problems arise: the structures of interest have to be segmented in the 4D data set and and the organ motion has to be estimated in the temporal image sequence. This paper presents a variational approach for simultaneous segmentation and registration applied to temporal image sequences. The proposed method assumes a known segmentation in one frame and then recovers nonlinear registration and segmentation in other frames by minimizing a cost function that combines intensity-based registration, level-set segmentation as well as prior shape and intensity knowledge. The purpose of the presented method is to estimate respiration induced organ motion in spatiotemporal CT image sequences and to segment a structure of interest simultaneously. A validation of the combined registration and segmentation approach is presented using low dose 4D CT data sets of the liver. The results demonstrate that the simultaneous solution of both problems improves the segmentation performance over a sequential application of the registration and segmentation steps.

The 3D-CSC is a general segmentation method for voxel images. One of its possible applications is the segmentation of MR images of the human head. We here propose a self-contained method consisting of preprocessing steps which remove common artifacts from the input image, a 3D-CSC segmentation which partitions the input image into gray value similar, spatially connected regions and a final classification of CSC segments into white matter, gray matter and non-brain. We evaluate our method using the brainweb dataset for which a ground truth is available.

We present a fully automatic method for segmentation of knee joint cartilage from fat suppressed MRI. The method first applies 3-D model-based segmentation technology, which allows to reliably segment the femur, patella, and tibia by iterative adaptation of the model according to image gradients. Thin plate spline interpolation is used in the next step to position deformable cartilage models for each of the three bones with reference to the segmented bone models. After initialization, the cartilage models are fine adjusted by automatic iterative adaptation to image data based on gray value gradients. The method has been validated on a collection of 8 (3 left, 5 right) fat suppressed datasets and demonstrated the sensitivity of 83±6% compared to manual segmentation on a per voxel basis as primary endpoint. Gross cartilage volume measurement yielded an average error of 9±7% as secondary endpoint. For cartilage being a thin structure, already small deviations in distance result in large errors on a per voxel basis, rendering the primary endpoint a hard criterion.

Cervical Intraepithelial Neoplasia (CIN) exhibits certain morphologic features that can be identified during a visual inspection exam. Immature and dysphasic cervical squamous epithelium turns white after application of acetic acid during the exam. The whitening process occurs visually over several minutes and subjectively discriminates between dysphasic and normal tissue. Digital imaging technologies allow us to assist the physician analyzing the acetic acid induced lesions (acetowhite region) in a fully automatic way. This paper reports a study designed to measure multiple parameters of the acetowhitening process from two images captured with a digital colposcope. One image is captured before the acetic acid application, and the other is captured after the acetic acid application. The spatial change of the acetowhitening is extracted using color and texture information in the post acetic acid image; the temporal change is extracted from the intensity and color changes between the post acetic acid and pre acetic acid images with an automatic alignment. The imaging and data analysis system has been evaluated with a total of 99 human subjects and demonstrate its potential to screening underserved women where access to skilled colposcopists is limited.

Automated image processing techniques may prove invaluable in the examination of real-time three-dimensional echocardiograms, by providing quantitative and objective measurements of functional parameters such as left ventricular (LV) volume and ejection fraction. In this study, we investigate the use of active appearance models (AAMs) for automatic detection of left ventricular endocardial contours. AAMs are especially useful in segmenting ultrasound images, due to their ability to model the typical LV appearance. However, since only a limited number of images is available for training, the model may be incapable of capturing the large variability in ultrasound image appearance. This may cause standard AAM matching procedures to fail if the model and image are significantly different. Recently, a Jacobian-tuning method for AAM matching was proposed, which allowed the model's training matrix to adapt to the new, unseen image. This may potentially result in a more robust matching. To compare both matching methods, AAMs were built with end-diastolic images from 54 patients. Larger capture ranges and higher accuracy were obtained when the new method was used. In conclusion, this method has great potential for segmentation in echocardiograms and will improve the assessment of LV functional parameters.

Colon cancer is one of the most frequent causes of death. CT colonography is a novel method for the detection of polyps and early cancer. The general principle of CT colonography includes a cathartic bowel preparation. The resulting discomfort for patients leads to limited patient acceptance and therefore to limited cancer detection rates. Reduced bowel preparation, techniques for stool tagging, and electronic cleansing, however, improve the acceptance rates. Hereby, the high density of oral contrast material highlights residual stool and can be digitally removed. Known subtraction methods cause artifacts: additional 3D objects are introduced and small bowel folds are perforated. We propose a new algorithm that is based on the 2^nd derivative of the image data using the Hessian matrix and the following principal axis transform to detect tiny folds which shall not be subtracted together with tagged stool found by a thresholding method. Since the stool is usually not homogenously tagged with contrast media a detection algorithm for island-like structures is incorporated. The interfaces of air-stool level and colon wall are detected by a 3-dimensional difference of Gaussian module. A 3-dimensional filter smoothes the transitions between removed stool and colon tissue. We evaluated the efficacy of the new algorithm with 10 patient data sets. The results showed no introduced artificial objects and no perforated folds. The artifacts at the air-stool and colon tissue-stool transitions are considerably reduced compared to those known from the literature.

We describe a method for segmenting airway trees from greyscale 3D images such as CT (Computed Tomography) scans. Our approach is based on topological analysis of sets obtained by thresholding from thick slices, i.e. sub-images consisting of a small number of consecutive slices. From each thick slice under consideration, we select all sets S obtained from that thick slice by thresholding that have simple enough topological structure. As the selection criterion, we use a simple algebraic condition involving the numbers of connected components in the intersection of the set S with every slice in the thick slice. The condition basically asserts that the intersections of S with each of the slices is small and attempts to limit the number of the branching points of S within the thick slice. The output 3D model of the airway tree is obtained as the largest connected component of the union of all selected sets, extracted from several overlapping thick slices. Experiments with a number of chest CT scans show that the method leads to promising results.

Image analysis for automated diagnosis of cervical cancer has attained high prominence in the last decade. Automated image analysis at all levels requires a basic segmentation of the region of interest (ROI) within a given image. The precision of the diagnosis is often reflected by the precision in detecting the initial region of interest, especially when some features outside the ROI mimic the ones within the same. Work described here discusses algorithms that are used to improve the cervical region of interest as a part of automated cervical image diagnosis. A vital visual aid in diagnosing cervical cancer is the aceto-whitening of the cervix after the application of acetic acid. Color and texture are used to segment acetowhite regions within the cervical ROI. Vaginal walls along with cottonswabs sometimes mimic these essential features leading to several false positives. Work presented here is focused towards detecting in-focus vaginal wall boundaries and then extrapolating them to exclude vaginal walls from the cervical ROI. In addition, discussed here is a marker-controlled watershed segmentation that is used to detect cottonswabs from the cervical ROI. A dataset comprising 50 high resolution images of the cervix acquired after 60 seconds of acetic acid application were used to test the algorithm. Out of the 50 images, 27 benefited from a new cervical ROI. Significant improvement in overall diagnosis was observed in these images as false positives caused by features outside the actual ROI mimicking acetowhite region were eliminated.