This PDF file contains the front matter associated with SPIE Proceedings Volume 7259, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and the Conference Committee listing.

Whole body CT scanning is a common diagnosis technique for discovering early signs of metastasis or for differential diagnosis. Automatic parsing and segmentation of multiple organs and semantic navigation inside the body can help the clinician in efficiently obtaining accurate diagnosis. However, dealing with the large amount of data of a full body scan is challenging and techniques are needed for the fast detection and segmentation of organs, e.g., heart, liver, kidneys, bladder, prostate, and spleen, and body landmarks, e.g., bronchial bifurcation, coccyx tip, sternum, lung tips. Solving the problem becomes even more challenging if partial body scans are used, where not all organs are present. We propose a new approach to this problem, in which a network of 1D and 3D landmarks is trained to quickly parse the 3D CT data and estimate which organs and landmarks are present as well as their most probable locations and boundaries. Using this approach, the segmentation of seven organs and detection of 19 body landmarks can be obtained in about 20 seconds with state-of-the-art accuracy and has been validated on 80 CT full or partial body scans.

Markov Random Fields (MRFs) provide a tractable means for incorporating contextual information into a Bayesian framework. This contextual information is modeled using multiple local conditional probability density functions (LCPDFs) which the MRF framework implicitly combines into a single joint probability density function (JPDF) that describes the entire system. However, only LCPDFs of certain functional forms are consistent, meaning they reconstitute a valid JPDF. These forms are specified by the Gibbs-Markov equivalence theorem which indicates that the JPDF, and hence the LCPDFs, should be representable as a product of potential functions (i.e. Gibbs distributions). Unfortunately, potential functions are mathematical abstractions that lack intuition; and consequently, constructing LCPDFs through their selection becomes an ad hoc procedure, usually resulting in generic and/or heuristic models. In this paper we demonstrate that under certain conditions the LCDPFs can be formulated in terms of quantities that are both meaningful and descriptive: probability distributions. Using probability distributions instead of potential functions enables us to construct consistent LCPDFs whose modeling capabilities are both more intuitive and expansive than typical MRF models. As an example, we compare the efficacy of our so-called probabilistic pairwise Markov models (PPMMs) to the prevalent Potts model by incorporating both into a novel computer aided diagnosis (CAD) system for detecting prostate cancer in whole-mount histological sections. Using the Potts model the CAD system is able to detection cancerous glands with a specificity of 0.82 and sensitivity of 0.71; its area under the receiver operator characteristic (AUC) curve is 0.83. If instead the PPMM model is employed the sensitivity (specificity is held fixed) and AUC increase to 0.77 and 0.87.

Traditional fuzzy clustering algorithms have been successfully applied in MR image segmentation for quantitative morphological analysis. However, the clustering results might be biased due to the variability of tissue intensities and anatomical structures. For example, clustering-based algorithms tend to over-segment white matter tissues of MR brain images. To solve this problem, we introduce a tissue probability map constrained clustering algorithm and apply it to serialMR brain image segmentation for longitudinal study of human brains. The tissue probability maps consist of segmentation priors obtained from a population and reflect the probability of different tissue types. More accurate image segmentation can be achieved by using these segmentation priors in the clustering algorithm. Experimental results of both simulated longitudinal MR brain data and the Alzheimer's Disease Neuroimaging Initiative (ADNI) data using the new serial image segmentation algorithm in the framework of CLASSIC show more accurate and robust longitudinal measures.

In this paper we present WERITAS, which is based in part on the traditional Active Shape Model (ASM) segmentation system. WERITAS generates multiple statistical texture features, and finds the optimal weighted average of those texture features by maximizing the correlation between the Euclidean distance to the ground truth and the Mahalanobis distance to the training data. The weighted average is used a multi-resolution segmentation system to more accurately detect the object border. A rigorous evaluation was performed on over 200 clinical images comprising of prostate images and breast images from 1.5 Tesla and 3 Tesla MRI machines via 6 distinct metrics. WERITAS was tested against a traditional multi-resolution ASM in addition to an ASM system which uses a plethora of random features to determine if the selection of features is improving the results rather than simply the use of multiple features. The results indicate that WERITAS outperforms all other methods to a high degree of statistical significance. For 1.5T prostate MRI images, the overlap from WERITAS is 83%, the overlap from the random features is 81%, and the overlap from the traditional ASM is only 66%. In addition, using 3T prostate MRI images, the overlap from WERITAS is 77%, the overlap from the random features is 54%, and the overlap from the traditional ASM is 59%, suggesting the usefulness of WERITAS. The only metrics in which WERITAS was outperformed did not hold any degree of statistical significance. WERITAS is a robust, efficient, and accurate segmentation system with a wide range of applications.

Magnetic resonance imaging (MRI) is currently the gold standard for left ventricle (LV) quantification. Detection of the LV in an MRI image is a prerequisite for functional measurement. However, due to the large variations in orientation, size, shape, and image intensity of the LV, automatic detection of the LV is still a challenging problem. In this paper, we propose to use marginal space learning (MSL) to exploit the recent advances in learning discriminative classifiers. Instead of learning a monolithic classifier directly in the five dimensional object pose space (two dimensions for position, one for orientation, and two for anisotropic scaling) as full space learning (FSL) does, we train three detectors, namely, the position detector, the position-orientation detector, and the position-orientation-scale detector. Comparative experiments show that MSL significantly outperforms FSL in both speed and accuracy. Additionally, we also detect several LV landmarks, such as the LV apex and two annulus points. If we combine the detected candidates from both the whole-object detector and landmark detectors, we can further improve the system robustness. A novel voting based strategy is devised to combine the detected candidates by all detectors. Experiments show component-based voting can reduce the detection outliers.

This paper presents a novel model based segmentation technique for quantification of Left Ventricular (LV) function from sparse single-beat 3D echocardiographic data acquired with a Fast Rotating Ultrasound (FRU) transducer. This transducer captures cardiac anatomy in a sparse set of radially sampled, curved cross sections within a single cardiac cycle. The method employs a 3D Active Shape Model of the Left Ventricle (LV) in combination with local appearance models as prior knowledge to steer the segmentation. A set of local appearance patches generate the model update points for fitting the model to the LV in the curved FRU cross-sections. Updates are then propagated over the dense 3D model mesh to overcome correspondence problems due to the data sparsity, whereas the 3D Active Shape Model serves to retain the plausibility of the generated shape. Leave-one-out cross validation was carried out on single-beat FRU data from 28 patients suffering from various cardiac pathologies. Error measurements against expert-annotated contours yielded an average point-to-point distance of around 3.8 ± 2.4 mm and point-to-surface distance of 2.0 ± 1.8 mm and average volume estimation error of around 9 ± 7%. Robustness tests with respect to different model initializations showed acceptable performance for initial positions within a range of 22 mm for displacement and 12° for orientation. This demonstrates that the method combines robustness with respect to initialization with an acceptable accuracy, while using sparse single-beat FRU data.

Disorders of the aortic valve represent a common cardiovascular disease and an important public-health problem worldwide. Pathological valves are currently determined from 2D images through elaborate qualitative evalu- ations and complex measurements, potentially inaccurate and tedious to acquire. This paper presents a novel diagnostic method, which identies diseased valves based on 3D geometrical models constructed from volumetric data. A parametric model, which includes relevant anatomic landmarks as well as the aortic root and lea ets, represents the morphology of the aortic valve. Recently developed robust segmentation methods are applied to estimate the patient specic model parameters from end-diastolic cardiac CT volumes. A discriminative distance function, learned from equivalence constraints in the product space of shape coordinates, determines the corresponding pathology class based on the shape information encoded by the model. Experiments on a heterogeneous set of 63 patients aected by various diseases demonstrated the performance of our method with 94% correctly classied valves.

This paper proposes a brain image registration algorithm, called RABBIT, which achieves fast and accurate image registration by using an intermediate template generated by a statistical shape deformation model during the image registration procedure. The statistical brain shape deformation information is learned by means of principal component analysis (PCA) from a set of training brain deformations, each of them linking a selected template to an individual brain sample. Using the statistical deformation information, the template image can be registered to a new individual image by optimizing a statistical deformation model with a small number of parameters, thus generating an intermediate template very close to the individual brain image. The remaining shape difference between the intermediate template and the individual brain is then minimized by a general registration algorithm, such as HAMMER. With the help of the intermediate template, the registration between the template and individual brain images can be achieved fast and with similar registration accuracy as HAMMER. The effectiveness of the RABBIT has been evaluated by using both simulated atrophy data and real brain images. The experimental results show that RABBIT can achieve over five times speedup, compared to HAMMER, without losing any registration accuracy or statistical power in detecting brain atrophy.

Estimation of the covariance matrix is a pivotal step in landmark based statistical shape analysis. For high dimensional representation of the shapes, often the number of available shape examples is far too small for ML covariance matrix is rank deficient and eigenvectors corresponding to the small eigenvalues. We take a Bayesian approach to the problem and show how the prior information can be used to estimate the covariance matrix from a small number of samples in a high dimensional shape space. The performance of the proposed method is evaluated in the context of reconstructions of high resolution vertebral boundary from an incomplete and lower dimensional representation. The algorithm performs better than the ML method, especially for small numbers of samples in the training set. The superiority of the proposed Bayesian approach was also observed when noisy incomplete lower dimensional representation of the vertebral boundary was used in the reconstruction algorithm. Moreover, unlike other commonly used approaches, e.g., regularization, the presented method does not depend heavily on the choice of the parameter values.

In this paper, we describe and compare methods for automatically identifying individual vertebrae in arbitrary CT images. The identification is an essential precondition for a subsequent model-based segmentation, which is used in a wide field of orthopedic, neurological, and oncological applications, e.g., spinal biopsies or the insertion of pedicle screws. Since adjacent vertebrae show similar characteristics, an automated labeling of the spine column is a very challenging task, especially if no surrounding reference structures can be taken into account. Furthermore, vertebra identification is complicated due to the fact that many images are bounded to a very limited field of view and may contain only few vertebrae. We propose and evaluate two methods for automatically labeling the spine column by evaluating similarities between given models and vertebral objects. In one method, object boundary information is taken into account by applying a Generalized Hough Transform (GHT) for each vertebral object. In the other method, appearance models containing mean gray value information are registered to each vertebral object using cross and local correlation as similarity measures for the optimization function. The GHT is advantageous in terms of computational performance but cuts back concerning the identification rate. A correct labeling of the vertebral column has been successfully performed on 93% of the test set consisting of 63 disparate input images using rigid image registration with local correlation as similarity measure.

Image segmentation methods may be classified into two categories: purely image based and model based. Each of these two classes has its own advantages and disadvantages. In this paper, we propose a novel synergistic combination of the image based graph-cut (GC) methods with the model based ASM methods to arrive at the GC-ASM method. GC-ASM effectively combines the rich statistical shape information embodied in ASM with the globally optimal delineation capability of the GC method. We propose a new GC cost function, which effectively integrates the specific image information with the ASM shape information. The ASM results are fully utilized to help GC in several ways: (1) For automatically selecting seeds to do GC segmentation, thus helping GC with object recognition; (2) For refining the parameters of the GC algorithm from the ASM result; (3) For bringing object shape information into the GC cost computation. (4) In turn, for using the cost of GC result to improve the ASM's object recognition process. The proposed methods are implemented to operate on 2D images and tested on a clinical abdominal CT data set. The results show: (1) GC-ASM becomes largely independent of initialization. (2) The number of landmarks can be reduced by a factor of 3 in GC-ASM over that in ASM. (3) The accuracy of segmentation via GC-ASM is considerably better than that of ASM.

The quantitative characterization of tissue probes as visualized by CT or MR is of great interest in many fields of medical image analysis. A proper quantification of the information content in such images can be realized by calculating well-suited texture measures, which are able to capture the main characteristics of the image structures under study. Using test images showing the complex trabecular structure of the inner bone of a healthy and osteoporotic patient we propose and apply a novel statistical framework, with which one can systematically assess the sensitivity of the chosen texture measures to higher order correlations (HOCs), i.e. correlations not being captured by linear methods like the power spectrum. To this end, so-called surrogate images are generated, in which the linear properties are preserved, while parts or all higher order correlations are wiped out. This is achieved by dedicated Fourier phase shuffling techniques. We compare three commonly used classes of texture measures, namely spherical Mexican hat wavelets (SMHW), Minkowski functionals (MF) and scaling indices (SIM). While the SMHW yield only very poor sensitivity to HOCs in both cases, the MF and SIM could detect the HOCs very well with significance up to S = 320σ (MF) and S = 150σ (SIM). The relative performance of the MF and SIM differed significantly for the healthy and osteoporotic bone. Thus, MF and SIM are preferable for a proper quantification of the bone structure. They depict complementary aspects of it and thus should both be used for characterising the trabecular bone.

We propose a method for retrieving similar fMRI statistical images given a query fMRI statistical image. Our method thresholds the voxels within those images and extracts spatially distinct regions from the voxels that remain. Each region is defined by a feature vector that contains the region centroid, the region area, the average activation value for all the voxels within that region, the variance of those activation values, the average distance of each voxel within that region to the region's centroid, and the variance of the voxel's distance to the region's centroid. The similarity between two images is obtained by the summed minimum distance of their constituent feature vectors. Results on a dataset of fMRI statistical images from experiments involving distinct cognitive tasks are shown.

Longitudinal imaging studies are essential to understanding the neural development of neuropsychiatric disorders, substance use disorders, and normal brain. Using appropriate image processing and statistical tools to analyze the imaging, behavioral, and clinical data is critical for optimally exploring and interpreting the findings from those imaging studies. However, the existing imaging processing and statistical methods for analyzing imaging longitudinal measures are primarily developed for cross-sectional neuroimaging studies. The simple use of these cross-sectional tools to longitudinal imaging studies will significantly decrease the statistical power of longitudinal studies in detecting subtle changes of imaging measures and the causal role of time-dependent covariate in disease process. The main objective of this paper is to develop longitudinal statistics toolbox, called LSTGEE, for the analysis of neuroimaging data from longitudinal studies. We develop generalized estimating equations for jointly modeling imaging measures with behavioral and clinical variables from longitudinal studies. We develop a test procedure based on a score test statistic and a resampling method to test linear hypotheses of unknown parameters, such as associations between brain structure and function and covariates of interest, such as IQ, age, gene, diagnostic groups, and severity of disease. We demonstrate the application of our statistical methods to the detection of the changes of the fractional anisotropy across time in a longitudinal neonate study. Particularly, our results demonstrate that the use of longitudinal statistics can dramatically increase the statistical power in detecting the changes of neuroimaging measures. The proposed approach can be applied to longitudinal data with multiple outcomes and accommodate incomplete and unbalanced data, i.e., subjects with different number of measurements.

In medical X-ray examinations, images suffer considerably from severe, signal-dependent noise as a result of the effort to keep applied doses as low as possible. This noise can be seen as an additive signal that degrades image quality and might disguise valuable content. Lost information has to be restored in a post-processing step. The crucial aspect of filtering medical images is preservation of edges and texture on the one hand and removing noise on the other hand. Classical smoothing filters, such as Gaussian or box filtering. are data-independent and equally blur the image content. State-of-the-art methods currently make use of local neighborhoods or global image statistics. However, exploiting global self-similarity within an image and inter-image similarity for subsequent frames of a sequence bears an unused potential for image restoration. We introduce a non-local filter with data-dependent response that closes the gap between local filtering and stochastic methods. The filter is based on the non-local means approach proposed by Buades¹ et al. and is similar to bilateral filtering. In order to apply this approach to medical data, we heavily reduce the computational costs incurred by the original approach. Thus it is possible to interactively enhance single frames or selected regions of interest within a sequence. The proposed filter is applicable for time-domain filtering without the need for accurate motion estimation. Hence it can be seen as a general solution for filtering 2D as well as 2D+t X-ray image data.

The knowledge of thresholding and gradient at different tissue interfaces is of paramount interest in image segmentation and other imaging methods and applications. Most thresholding and gradient selection methods primarily focus on image histograms and therefore, fail to harness the information generated by intensity patterns in an image. We present a new thresholding and gradient optimization method which accounts for spatial arrangement of intensities forming different objects in an image. Specifically, we recognize object class uncertainty, a histogram-based feature, and formulate an energy function based on its correlation with image gradients that characterizes the objects and shapes in a given image. Finally, this energy function is used to determine optimum thresholds and gradients for various tissue interfaces. The underlying theory behind the method is that objects manifest themselves with fuzzy boundaries in an acquired image and that, in a probabilistic sense; intensities with high class uncertainty are associated with high image gradients generally indicating object/tissue interfaces. The new method simultaneously determines optimum values for both thresholds and gradient parameters at different object/tissue interfaces. The method has been applied on several 2D and 3D medical image data sets and it has successfully determined both thresholds and gradients for different tissue interfaces even when some of the thresholds are almost impossible to locate in their histograms. The accuracy and reproducibility of the method has been examined using 3D multi-row detector computed tomography images of two cadaveric ankles each scanned thrice with repositioning the specimen between two scans.

Cumulative residual entropy (CRE)^1,2 has recently been advocated as an alternative to differential entropy for describing the complexity of an image. CRE has been used to construct an alternate form of mutual information (MI),^3,4 called symmetric cumulative mutual information (SCMI)⁵ or cross-CRE (CCRE).⁶ This alternate form of MI has exhibited superior performance to traditional MI in a variety of ways.6 However, like traditional MI, SCMI suffers from sensitivity to the changing size of the overlap between images over the course of registration. Alternative similarity measures based on differential entropy, such as normalized mutual information (NMI),⁷ entropy correlation coefficient (ECC)^8,9 and modified mutual information (M-MI),¹⁰ have been shown to exhibit superior performance to MI with respect to the overlap sensitivity problem. In this paper, we show how CRE can be used to compute versions of NMI, ECC, and M-MI that we call the normalized cumulative mutual information (NCMI), cumulative residual entropy correlation coefficient (CRECC), and modified symmetric cumulative mutual information (M-SCMI). We use publicly available CT, PET, and MR brain images* with known ground truth transformations to evaluate the performance of these CRE-based similarity measures for rigid multimodal registration. Results show that the proposed similarity measures provide a statistically significant improvement in target registration error (TRE) over SCMI.

Functional MRI (fMRI) time-series studies are plagued by varying degrees of subject head motion. Faithful head motion correction is essential to accurately detect brain activation using statistical analyses of these time-series. Mutual information (MI) based slice-to-volume (SV) registration is used for motion estimation when the rate of change of head position is large. SV registration accounts for head motion between slice acquisitions by estimating an independent rigid transformation for each slice in the time-series. Consequently each MI optimization uses intensity counts from a single time-series slice, making the algorithm susceptible to noise for low complexity endslices (i.e., slices near the top of the head scans). This work focuses on improving the accuracy of MI-based SV registration of end-slices by using joint probability density priors derived from registered high complexity centerslices (i.e., slices near the middle of the head scans). Results show that the use of such priors can significantly improve SV registration accuracy.

A novel method for quality assessment in medical image registration is presented. It is evaluated on 24 follow-up CT scan pairs of the lung. Based on a reference standard of manually matched landmarks we established a pattern recognition approach for detection of local registration errors. To capture characteristics of these misalignments a set of intensity, entropy and deformation related features was employed. Feature selection was conducted and a kNN classifier was trained and evaluated on a subset of landmarks. Registration errors larger than 2 mm were classified with a sensitivity of 88% and specificity of 94%.

Non-rigid registration techniques are commonly used in medical image analysis. However these techniques are often time consuming. Graphics Processing Unit (GPU) execution appears to be a good way to decrease computation time significantly. However for an efficient implementation on GPU, an algorithm must be data parallel. In this paper we compare the analytical calculation of the gradient of Normalised Mutual Information with an approximation better suited to parallel implementation. Both gradient approaches have been implemented using a Free-Form Deformation framework based on cubic B-Splines and including a smoothness constraint. We applied this technique to recover realistic deformation fields generated from 65 3D-T1 images. The recovered fields using both gradients and the ground truth were compared. We demonstrated that the approximated gradient performed similarly to the analytical gradient but with a greatly reduced computation time when both approaches are implemented on the CPU. The implementation of the approximated gradient on the GPU leads to a computation time of 3 to 4 minutes when registering 190 × 200 × 124 voxel images with a grid including 57 × 61 × 61 control points.

We present a new non-uniform adaptive sampling method for the estimation of mutual information in multi-modal image registration. The method uses the Fast Discrete Curvelet Transform to identify regions along anatomical curves on which the mutual information is computed. Its main advantages of over other non-uniform sampling schemes are that it captures the most informative regions, that it is invariant to feature shapes, orientations, and sizes, that it is efficient, and that it yields accurate results. Extensive evaluation on 20 validated clinical brain CT images to Proton Density (PD) and T1 and T2-weighted MRI images from the public RIRE database show the effectiveness of our method. Rigid registration accuracy measured at 10 clinical targets and compared to ground truth measurements yield a mean target registration error of 0.68mm(std=0.4mm) for CT-PD and 0.82mm(std=0.43mm) for CT-T2. This is 0.3mm (1mm) more accurate in the average (worst) case than five existing sampling methods. Our method has the lowest registration errors recorded to date for the registration of CT-PD and CT-T2 images in the RIRE website when compared to methods that were tested on at least three patient datasets.

Group-wise registration has been proposed recently for consistent registration of all images in the same dataset. Since all images need to be registered simultaneously with lots of deformation parameters to be optimized, the number of images that the current group-wise registration methods can handle is limited due to the capability of CPU and physical memory in a general computer. To overcome this limitation, we present a hierarchical group-wise registration method for feasible registration of large image dataset. Our basic idea is to decompose the large-scale group-wise registration problem into a series of small-scale registration problems, each of which can be easily solved. In particular, we use a novel affinity propagation method to hierarchically cluster a group of images into a pyramid of classes. Then, images in the same class are group-wisely registered to their own center image. The center images of different classes are further group-wisely registered from the lower level to the upper level of the pyramid. A final atlas for the whole image dataset is thus synthesized when the registration process reaches the top of the pyramid. By applying this hierarchical image clustering and atlas synthesis strategy, we can efficiently and effectively perform group-wise registration to a large image dataset and map each image into the atlas space. More importantly, experimental results on both real and simulated data also confirm that the proposed method can achieve more robust and accurate registration than the conventional group-wise registration algorithms.

Progression measurement of emphysema is required to evaluate the health condition of a patient and the effect of drugs. To locally estimate progression we use image registration, which allows for volume correction using the determinant of the Jacobian of the transformation. We introduce an adaptation of the so-called sponge model that circumvents its constant-mass assumption. Preliminary results from CT scans of a lung phantom and from CT data sets of three patients suggest that image registration may be a suitable method to locally estimate emphysema progression.

This paper describes a novel method for registering multimodal breast images. The method is based on guiding initial alignment by a 3D statistical deformation model (SDM) followed by a standard non-rigid registration method for fine alignment. The method was applied to the problem of compensating for large breast compressions, namely registering magnetic resonance (MR) images to tomosynthesis images and X-ray mammograms. The SDM was based on simulating plausible breast compressions for a population of 20 subjects via finite element models created from segmented 3D MR breast images. Leave-one-out tests on simulated data showed that using SDM guided registration rather than affine registration for the initial alignment led on average to lower mean registration errors, namely 3.2 mm versus 4.2 mm for MR to tomosynthesis images (17.1 mm initially) and 5.0 mm versus 6.2 mm for MR to X-ray mammograms (15.0 mm initially).

Although useful for the detection of breast cancers, conventional imaging methods, including mammography and ultrasonography, do not provide adequate information regarding response to therapy. Dynamic contrast enhanced MRI (DCE-MRI) has emerged as a promising technique to provide relevant information on tumor status. Consequently, accurate longitudinal registration of breast MR images is critical for the comparison of changes induced by treatment at the voxel level. In this study, a nonrigid registration algorithm is proposed to allow for longitudinal registration of breast MR images obtained throughout the course of treatment. We accomplish this by modifying the adaptive bases algorithm (ABA) through adding a tumor volume preserving constraint in the cost function. The registration results demonstrate the proposed algorithm can successfully register the longitudinal breast MR images and permit analysis of the parameter maps. We also propose a novel validation method to evaluate the proposed registration algorithm quantitatively. These validations also demonstrate that the proposed algorithm constrains tumor deformation well and performs better than the unconstrained ABA algorithm.

Coronary CT angiography (CTA) with multi-slice helical scanners is becoming the integral part of major diagnostic pathways for coronary artery disease. In addition, coronary CTA has demonstrated substantial potential in quantitative coronary plaque characterization. If serial comparisons of plaque progression or regression are to be made, accurate 3D volume registration of these volumes would be particularly useful. In this work, we propose a coronary CTA volume registration of the paired coronary CTA scans using feature-based non-rigid volume registration. We achieve this with a combined registration strategy, which uses the global rigid registration as an initialization, followed by local registration using non-rigid volume registration with a volume preserving constraint. We exploit the extracted coronary trees to help localize and emphasize the region of interest as unnecessary regions hinder registration process, which results in wrong registration result. The extracted binary masks of each coronary tree may not be the same due to initial segmentation errors, which could lead to subsequent bias in the registration process. Therefore we utilize a blur mask which is generated by convolving the Gaussian function with the binary coronary tree mask to include the neighboring vessel region into account. A volume preserving constraint is imposed so that the total volume of the binary mask before and after co-registration remains constant. To validate the proposed method, we perform experiments with data from 3 patients with available serial CT scans (6 scans in total) and measure the distance of anatomical landmarks between the registered serial scans of the same patient.

Morphological changes of Doppler ultrasound images are an important source of information for diagnosis of cardiovascular diseases. Quantification of these flow profiles requires segmentation of the ultrasound images. In this article, we propose a new model-based method for segmentation of (aortic outflow) velocity profiles. The method is based on a procedure for registration using a geometric transformation specifically designed for matching Doppler ultrasound profiles. After manual segmentation of a model image, the model image is temporarily registered to a new image using two manually defined points in time. Next, a non-rigid registration was carried out in the velocity direction. As a similarfity measure normalized mutual information is used, while optimization is performed by a genetic algorithm. The registration method is experimentally validated using an in-silico image phantom, and showed an accuracy of 5.4%. The model based on segmentation is evaluated in a seris of aortic outflow Doppler ultrasound images from 30 normal volunteers. Comparing the automated method to the manual delineation by an expert cardiologist the method proved accurate to 6.6%. The experimental results confirm the accuracy of the approach and shows that the method can be used for the segmentation of the clinically obtained aortic outflow velocity profiles.

Respiration-induced organ motion can limit the accuracy required for many clinical applications working on the thorax or upper abdomen. One approach to reduce the uncertainty of organ location caused by respiration is to use prior knowledge of breathing motion. In this work, we deal with the extraction and modeling of lung motion fields based on free-breathing 4D-CT data sets of 36 patients. Since data was acquired for radiotherapy planning, images of the same patient were available over different weeks of treatment. Motion field extraction is performed using an iterative shape-constrained deformable model approach. From the extracted motion fields, intra- and inter-subject motion models are built and adapted in a leave-one-out test. The created models capture the motion of corresponding landmarks over the breathing cycle. Model adaptation is then performed by examplarily assuming the diaphragm motion to be known. Although, respiratory motion shows a repetitive character, it is known that patients' variability in breathing pattern impedes motion estimation. However, with the created motion models, we obtained a mean error between the phases of maximal distance of 3.4 mm for the intra-patient and 4.2 mm for the inter-patient study when assuming the diaphragm motion to be known.

Spatiotemporal image data allow analyzing respiratory dynamics and its impact on radiation therapy. A key feature within this field of research is the process of lung motion field estimation. For a multitude of applications feasible and "realistic" motion field estimates are required. Widely non-linear registration methods are applied to estimate motion fields; in this case physiology is not taken into account. Using Finite Element Methods we implemented a biophysical approach to model respiratory lung motion starting with the physiology of breathing. Resulting motion models are compared to motion field estimates of a non-linear non-parametric intensity-based registration approach. Additionally, we extended the registration approach to cope with discontinuities in pleura and chest wall motion as motivated by the biophysical model. Accuracy of the different modeling approaches is evaluated using a total of 800 user-defined landmarks in 4D(=3D+t) CT data of 10 lung tumor patients (between 70 and 90 landmarks each patient). Mean registration residuals (= difference between landmark motion as predicted model-based and as observed by an expert) are 3.2±2.0 mm (biophysical model), 3.4±2.4 mm (registration of segmented lung data), 2.1±2.3 mm (registration of CT data), and 1.6±1.3 mm (extended registration of CT data); intraobserver variability of landmark identification is 0.9±0.8 mm, mean landmark motion is 6.8±5.4 mm. Thus, prediction accuracy is higher for non-linear registration of the CT data, but it is shown that explicit modeling of boundary conditions motivated by the physiology of breathing and the biophysical modeling approach, respectively, improves registration accuracy significantly.

Harmonic phase (HARP) motion analysis is widely used in the analysis of tagged magnetic resonance images of the heart. HARP motion tracking can yield gross errors, however, when there is a large amount of motion between successive time frames. Methods that use spatial continuity of motion - so-called refinement methods - have previously been reported to reduce these errors. This paper describes a new refinement method based on shortest-path computations. The method uses a graph representation of the image and seeks an optimal tracking order from a specified seed to each point in the image by solving a single source shortest path problem. This minimizes the potential for path dependent solutions which are found in other refinement methods. Experiments on cardiac motion tracking shows that the proposed method can track the whole tissue more robustly and is also computationally efficient.

For obtaining quantitative and objective functional parameters from three-dimensional (3D) echocardiographic sequences, automated segmentation methods may be preferable to cumbersome manual delineation of 3D borders. In this study, a novel optical-flow based tracking method is proposed for propagating 3D endocardial contours of the left ventricle throughout the cardiac cycle. To take full advantage of the time-continuous nature of cardiac motion, a statistical motion model was explicitly embedded in the optical flow solution. The cardiac motion was modeled as frame-to-frame affine transforms, which were extracted using Procrustes analysis on a set of training contours. Principal component analysis was applied to obtain a compact model of cardiac motion throughout the whole cardiac cycle. The parameters of this model were resolved in an optical flow manner, via spatial and temporal gradients in image intensity. The algorithm was tested on 36 noncontrast and 28 contrast enhanced 3D echocardiographic sequences in a leave-one-out manner. Good results were obtained using a combination of the proposed motion-guided method and a purely data-driven optical flow approach. The improvement was particularly noticeable in areas where the LV wall was obscured by image artifacts. In conclusion, the results show the applicability of the proposed method in clinical quality echocardiograms.

In this paper, we present a novel approach for coronary artery motion modeling from cardiac Computed Tomography( CT) images. The aim of this work is to develop a 4D motion model of the coronaries for image guidance in robotic-assisted totally endoscopic coronary artery bypass (TECAB) surgery. To utilize the pre-operative cardiac images to guide the minimally invasive surgery, it is essential to have a 4D cardiac motion model to be registered with the stereo endoscopic images acquired intraoperatively using the da Vinci robotic system. In this paper, we are investigating the extraction of the coronary arteries and the modelling of their motion from a dynamic sequence of cardiac CT. We use a multi-scale vesselness filter to enhance vessels in the cardiac CT images. The centerlines of the arteries are extracted using a ridge traversal algorithm. Using this method the coronaries can be extracted in near real-time as only local information is used in vessel tracking. To compute the deformation of the coronaries due to cardiac motion, the motion is extracted from a dynamic sequence of cardiac CT. Each timeframe in this sequence is registered to the end-diastole timeframe of the sequence using a non-rigid registration algorithm based on free-form deformations. Once the images have been registered a dynamic motion model of the coronaries can be obtained by applying the computed free-form deformations to the extracted coronary arteries. To validate the accuracy of the motion model we compare the actual position of the coronaries in each time frame with the predicted position of the coronaries as estimated from the non-rigid registration. We expect that this motion model of coronaries can facilitate the planning of TECAB surgery, and through the registration with real-time endoscopic video images it can reduce the conversion rate from TECAB to conventional procedures.

Digital subtraction angiography (DSA) is a well-known technique for improving the visibility and perceptibility of blood vessels in the human body. Coronary DSA extends conventional DSA to dynamic 2D fluoroscopic sequences of coronary arteries which are subject to respiratory and cardiac motion. Effective motion compensation is the main challenge for coronary DSA. Without a proper treatment, both breathing and heart motion can cause unpleasant artifacts in coronary subtraction images, jeopardizing the clinical value of coronary DSA. In this paper, we present an effective method to separate the dynamic layer of background structures from a fluoroscopic sequence of the heart, leaving a clean layer of moving coronary arteries. Our method combines the techniques of learning-based vessel detection and robust motion estimation to achieve reliable motion compensation for coronary sequences. Encouraging results have been achieved on clinically acquired coronary sequences, where the proposed method considerably improves the visibility and perceptibility of coronary arteries undergoing breathing and cardiac movement. Perceptibility improvement is significant especially for very thin vessels. The potential clinical benefit is expected in the context of obese patients and deep angulation, as well as in the reduction of contrast dose in normal size patients.

3D CT Angiography (CTA) scans are currently used to assess the cerebral arteries. An additional 4D CT Perfusion (CTP) scan is often acquired to determine perfusion parameters in the cerebral parenchyma. We propose a method to extract a three dimensional volume showing either the arteries (arteriogram) or the veins (venogram) from the 4D CTP scan. This would allow cerebrovascular assessment using the CTP scan and obviate the need for acquiring an additional CTA scan. Preprocessing steps consist of registration of the time volumes of the CTP scan using rigid registration and masking out extracranial structures, bone and air. Next a 3D volume is extracted containing the vessels (vascular volume) by using the absolute area under the first derivative curve in time. To segment the arteries and veins we use the time to peak of the contrast enhancement curve combined with region growing within a rough vessel segmentation. Finally the artery/vein segmentation is used to suppress either the veins or the arteries in the vascular volume to construct the arteriogram and venogram. To evaluate the method, 11 arteriograms and venograms were visually inspected by an expert observer, with special attention to the important cerebral arteries (Circle of Willis) and veins (straight and transverse sinus). Results show that the proposed method is effective in extracting the major cerebral arteries and veins from CTP scans.

Distinguishing arterial and venous trees in pulmonary multiple-detector X-ray computed tomography (MDCT) images (contrast-enhanced or unenhanced) is a critical first step in the quantification of vascular geometry for purposes of determining, for instance, pulmonary hypertension, using vascular dimensions as a comparator for assessment of airway size, detection of pulmonary emboli and more. Here, a novel method is reported for separating arteries and veins in MDCT pulmonary images. Arteries and veins are modeled as two iso-intensity objects closely entwined with each other at different locations at various scales. The method starts with two sets of seeds -- one for arteries and another for veins. Initialized with seeds, arteries and veins grow iteratively while maintaining their spatial separation and eventually forming two disjoint objects at convergence. The method combines fuzzy distance transform, a morphologic feature, with a topologic connectivity property to iteratively separate finer and finer details starting at a large scale and progressing towards smaller scales. The method has been validated in mathematically generated tubular objects with different levels of fuzziness, scale and noise. Also, it has been successfully applied to clinical CT pulmonary data. The accuracy of the method has been quantitatively evaluated by comparing its results with manual outlining. For arteries, the method has yielded correctness of 81.7% at the cost of 6.7% false positives and 11.6% false negatives. Our method is very promising for automated separation of arteries and veins in MDCT pulmonary images even when there is no mark of intensity variation at conjoining locations.

The segmentation of the hepatic vascular tree in computed tomography (CT) images is important for many applications such as surgical planning of oncological resections and living liver donations. In surgical planning, vessel segmentation is often used as basis to support the surgeon in the decision about the location of the cut to be performed and the extent of the liver to be removed, respectively. We present a novel approach to hepatic vessel segmentation that can be divided into two stages. First, we detect and delineate the core vessel components efficiently with a high specificity. Second, smaller vessel branches are segmented by a robust vessel tracking technique based on a medialness filter response, which starts from the terminal points of the previously segmented vessels. Specifically, in the first phase major vessels are segmented using the globally optimal graphcuts algorithm in combination with foreground and background seed detection, while the computationally more demanding tracking approach needs to be applied only locally in areas of smaller vessels within the second stage. The method has been evaluated on contrast-enhanced liver CT scans from clinical routine showing promising results. In addition to the fully-automatic instance of this method, the vessel tracking technique can also be used to easily add missing branches/sub-trees to an already existing segmentation result by adding single seed-points.

We present a novel method for lung vessel tree segmentation. The method combines image information and a high-level physiological model, stating that the vasculature is organized such that the whole organ is perfused using minimal effort. The method consists of three consecutive steps. First, a limited set of possible bifurcation locations is determined. Subsequently, individual vessel segments of varying diameters are constructed between each two bifurcation locations. This way, a graph is constructed consisting of each bifurcation location candidate as vertices and vessel segments as edges. Finally, the overall vessel tree is found by selecting the subset of these segments that perfuses the whole organ, while minimizing an energy function. This energy function contains a data term, a volume term and a bifurcation term. The data term measures how well the selected vessel segments fit to the image data, the volume term measures the total amount of blood in the vasculature, and the bifurcation term models the physiological fit of the diameters of the in- and outgoing vessels in each bifurcation. The selection of the optimal subset of vessel segments into a single vessel tree is an NP-hard combinatorial optimization problem that is solved here with an ant colony optimization approach. The bifurcation detection as well as the segmentation method have been validated on lung CT images with manually segmented arteries and veins.

We present a novel method for incorporating both edge and regional image information in a 3-D graph-theoretic approach for globally optimal surface segmentation. The energy functional takes a ratio form of the "onsurface" cost and the "in-region" cost. We thus introduce an optimal surface segmentation model allowing regional information such as volume, homogeneity and texture to be included with boundary information such as intensity gradients. Compared to the linear combination as in the standard active contour energies, this ratioform energy is parameter free with no bias toward either a large or small region. Our method is the first attempt to use a ratio-form energy functional in graph search framework for high dimensional image segmentation, which delivers a globally optimal solution in polynomial time. The globally optimal surface can be achieved by solving a parametric maximum flow problem in the time complexity of computing a single maximum flow. Our new approach is applied to the aorta segmentation of 15 3-D MR aortic images from 15 subjects. Compared to an expert-defined independent standard, the overall mean unsigned surface positioning error was 0.76± 0.88 voxels. Our experiments showed that the incorporation of the regional information was effective to alleviate the interference of adjacent objects.

Tridirectional Phase-Contrast (PC)-MRI sequences provide spatially and temporally resolved measurements of blood flow velocity vectors in the human body. Analyzing flow conditions based on these datasets requires prior segmentation of the vessels of interest. In view of decreased quality of morphology images in PC-MRI sequences, the flow data provides valuable information to support reliable segmentation. This work presents a semi-automatic approach for segmenting the large arteries utilizing both morphology and flow information. It consists of two parts, the extraction of a simplified vessel model based on vessel centerlines and diameters, and a following refinement of the resulting surface for each time frame. Vessel centerlines and diameters are extracted using an offset adaptive medialness function that estimates a voxel's likelihood of belonging to a vessel centerline. The resulting centerline model is manually post-processed to select the appropriate centerlines and link possible gaps. The surface described by the final centerline model is used to initialize a 3D level set segmentation of each time frame. Deformation velocities that depend on both morphology and flow information are proposed and a new approach to account for the curved shape of vessels is introduced. The described segmentation system has been successfully applied on a total of 22 datasets of the thoracic aorta and the pulmonary arteries. Resulting segmentations have been assessed by an expert radiologist and were considered to be very satisfactory.

Segmenting objects with complicated topologies in 3D images is a challenging problem in medical image processing, especially for objects with multiple interrelated surfaces. In this paper, we extend a graph search based technique to simultaneously identifying multiple interrelated surfaces for objects that have complex topologies (e.g., with tree-like structures) in 3D. We first perform a pre-segmentation on the input image to obtain basic information of the objects' topologies. Based on the initial pre-segmentation, the original image is resampled along judiciously determined directions to produce a set of vectors of voxels (called voxel columns). The resampling process utilizes medial axes to ensure that voxel columns of appropriate lengths are used to capture the sought object surfaces. Then a geometric graph is constructed whose edges connect voxels in the resampled voxel columns and enforce the smoothness constraint and separation constraint on the sought surfaces. Validation of our algorithm was performed on the segmentation of airway trees and lung vascular trees in human in-vivo CT scans. Cost functions with directional information are applied to distinguish the airway inner wall and outer wall. We succeed in extracting the outer airway wall and optimizing the location of the inner wall in all cases, while the vascular trees are optimized as well. Comparing with the pre-segmentation results, our approach captures the wall surfaces more accurately, especially across bifurcations. The statistical evaluation on a double wall phantom derived from in-vivo CT images yields highly accurate results of the wall thickness measurement on the whole tree (with mean unsigned error 0.16 ± 0.16mm).

In recent years, radiation therapy has become the preferred treatment for many types of head and neck tumors. To minimize side effects, radiation beams are planned pre-operatively to avoid over-radiation of vital structures, such as the optic nerves and chiasm, which are essential to the visual process. To plan the procedure, these structures must be identified using CT/MR imagery. Currently, a radiation oncologist must manually segment the structures, which is both inefficient and ineffective. Clearly an automated approach could be beneficial to the planning process. The problem is difficult due to the shape variability and low image contrast of the structures, and several attempts at automatic localization have been reported with marginal results. In this work we present a novel method for localizing the optic nerves and chiasm in CT/MR volumes using the atlas-navigated optimal medial axis and deformable-model algorithm (NOMAD). NOMAD uses a statistical model and image registration to provide a priori local intensity and shape information to both a medial axis extraction procedure and a deformable-model, which deforms the medial axis and completes the segmentation process. This approach achieves mean dice coefficients greater than 0.8 for both the optic nerves and the chiasm when compared to manual segmentations over ten test cases. By comparing quantitative results with existing techniques it can be seen that this method produces more accurate results.

Recent advances in MR and image analysis allow for reconstruction of high-resolution 3D images from clinical in utero scans of the human fetal brain. Automated segmentation of tissue types from MR images (MRI) is a key step in the quantitative analysis of brain development. Conventional atlas-based methods for adult brain segmentation are limited in their ability to accurately delineate complex structures of developing tissues from fetal MRI. In this paper, we formulate a novel geometric representation of the fetal brain aimed at capturing the laminar structure of developing anatomy. The proposed model uses a depth-based encoding of tissue occurrence within the fetal brain and provides an additional anatomical constraint in a form of a laminar prior that can be incorporated into conventional atlas-based EM segmentation. Validation experiments are performed using clinical in utero scans of 5 fetal subjects at gestational ages ranging from 20.5 to 22.5 weeks. Experimental results are evaluated against reference manual segmentations and quantified in terms of Dice similarity coefficient (DSC). The study demonstrates that the use of laminar depth-encoded tissue priors improves both the overall accuracy and precision of fetal brain segmentation. Particular refinement is observed in regions of the parietal and occipital lobes where the DSC index is improved from 0.81 to 0.82 for cortical grey matter, from 0.71 to 0.73 for the germinal matrix, and from 0.81 to 0.87 for white matter.

Accurate definition of the central-chest vasculature from three-dimensional (3D) multi-detector CT (MDCT) images is important for pulmonary applications. For instance, the aorta and pulmonary artery help in automatic definition of the Mountain lymph-node stations for lung-cancer staging. This work presents a system for defining major vascular structures in the central chest. The system provides automatic methods for extracting the aorta and pulmonary artery and semi-automatic methods for extracting the other major central chest arteries/veins, such as the superior vena cava and azygos vein. Automatic aorta and pulmonary artery extraction are performed by model fitting and selection. The system also extracts certain vascular structure information to validate outputs. A semi-automatic method extracts vasculature by finding the medial axes between provided important sites. Results of the system are applied to lymph-node station definition and guidance of bronchoscopic biopsy.

Purpose: By incorporating high-level shape priors, atlas-based segmentation has achieved tremendous success in the area of medical image analysis. However, the effect of various kinds of atlases, e.g., average shape model, example-based multi-atlas, has not been fully explored. In this study, we aim to generate different atlases and compare their performance in segmentation. Methods: We compare segmentation performance using parametric deformable model with four different atlases, including 1) a single atlas, i.e., average shape model (SAS); 2) example-based multi-atlas (EMA); 3) cluster-based average shape models (CAS); 4) cluster-based statistical shape models (average shape + principal shape variation modes)(CSS). CAS and CSS are novel atlases constructed by shape clustering. For comparison purpose, we also use PDM without atlas (NOA) as a benchmark method. Experiments: The experiment is carried on liver segmentation from whole-body CT images. Atlases are constructed by 39 manually delineated liver surfaces. 11 CT scans with ground truth are used as testing data set. Segmentation accuracy using different atlases are compared. Conclusion: Compared with segmentation without atlas, all of the four atlas-based image segmentation methods achieve better results. Multi-atlas based segmentation behaves better than single-atlas based segmentation. CAS exhibit superior performance to all other methods.

A robust model for the automatic segmentation of human brain images into anatomically defined regions across the human lifespan would be highly desirable, but such structural segmentations of brain MRI are challenging due to age-related changes. We have developed a new method, based on established algorithms for automatic segmentation of young adults' brains. We used prior information from 30 anatomical atlases, which had been manually segmented into 83 anatomical structures. Target MRIs came from 80 subjects (~12 individuals/decade) from 20 to 90 years, with equal numbers of men, women; data from two different scanners (1.5T, 3T), using the IXI database. Each of the adult atlases was registered to each target MR image. By using additional information from segmentation into tissue classes (GM, WM and CSF) to initialise the warping based on label consistency similarity before feeding this into the previous normalised mutual information non-rigid registration, the registration became robust enough to accommodate atrophy and ventricular enlargement with age. The final segmentation was obtained by combination of the 30 propagated atlases using decision fusion. Kernel smoothing was used for modelling the structural volume changes with aging. Example linear correlation coefficients with age were, for lateral ventricular volume, r_male=0.76, r_female=0.58 and, for hippocampal volume, r_male=-0.6, r_female=-0.4 (allρ<0.01).

Diffusion tensor image (DTI) is a powerful tool for quantitatively assessing the integrity of anatomical connectivity in white matter in clinical populations. The prevalent methods for group-level analysis of DTI are statistical analyses of invariant measures (e.g., fractional anisotropy) and principal directions across groups. The invariant measures and principal directions, however, do not capture all information in full diffusion tensor, which can decrease the statistical power of DTI in detecting subtle changes of white matters. Thus, it is very desirable to develop new statistical methods for analyzing full diffusion tensors. In this paper, we develop a set of toolbox, called RADTI, for the analysis of the full diffusion tensors as responses and establish their association with a set of covariates. The key idea is to use the recent development of log-Euclidean metric and then transform diffusion tensors in a nonlinear space into their matrix logarithms in a Euclidean space. Our regression model is a semiparametric model, which avoids any specific parametric assumptions. We develop an estimation procedure and a test procedure based on score statistics and a resampling method to simultaneously assess the statistical significance of linear hypotheses across a large region of interest. Monte Carlo simulations are used to examine the finite sample performance of the test procedure for controlling the family-wise error rate. We apply our methods to the detection of statistical significance of diagnostic and age effects on the integrity of white matter in a diffusion tensor study of human immunodeficiency virus.

Magnetic resonance diffusion tensor imaging (DTI) is being widely used to reconstruct brain white matter (WM) fiber tracts. For further characterization of the tracts, the fibers with similar courses often need to be grouped into a fiber bundle that corresponds to certain underlying WM anatomic structure. In addition, the alignments of fibers from different studies are often desirable for bundle comparisons and group analysis. In this work, a novel registration algorithm based on fiber-to-bundle matching was proposed to address the above two needs. Using an Expectation Maximization (EM) algorithm, the proposed method is capable of estimating a Thin-Plate- Spline transformation that optimally aligns whole-brain target fiber sets with a reference bundle model. Based on the resulting transformations, the fibers from different target datasets can all be warped into the reference coordinate system for comparisons and group analysis. The fibers can be further automatically labeled according to their similarity to the reference model. The algorithm was evaluated with eight human brain DTI data volumes acquired in vivo at 3T. After registration, the warped target bundles exhibit good similarity to the reference bundles. Quantitative experiments further demonstrated that the detected target bundles agree with ground truth obtained by manual segmentation at a sub-voxel accuracy.

This paper presents a segmentation technique for diffusion tensor imaging (DTI). This technique is based on a tensorial morphological gradient (TMG), defined as the maximum dissimilarity over the neighborhood. Once this gradient is computed, the tensorial segmentation problem becomes an scalar one, which can be solved by conventional techniques, such as watershed transform and thresholding. Similarity functions, namely the dot product, the tensorial dot product, the J-divergence and the Frobenius norm, were compared, in order to understand their differences regarding the measurement of tensor dissimilarities. The study showed that the dot product and the tensorial dot product turned out to be inappropriate for computation of the TMG, while the Frobenius norm and the J-divergence were both capable of measuring tensor dissimilarities, despite the distortion of Frobenius norm, since it is not an affine invariant measure. In order to validate the TMG as a solution for DTI segmentation, its computation was performed using distinct similarity measures and structuring elements. TMG results were also compared to fractional anisotropy. Finally, synthetic and real DTI were used in the method validation. Experiments showed that the TMG enables the segmentation of DTI by watershed transform or by a simple choice of a threshold. The strength of the proposed segmentation method is its simplicity and robustness, consequences of TMG computation. It enables the use, not only of well-known algorithms and tools from the mathematical morphology, but also of any other segmentation method to segment DTI, since TMG computation transforms tensorial images in scalar ones.

This paper proposes a new pairwise registration algorithm called Large Diffeomorphic Free-Form Deformation (LDFFD). In the LDFFD algorithm, the diffeomorphic mapping is represented as a composition of Free-Form Deformation (FFD) transformations at each time step, all time steps being jointly optimized. The fact that the transformation at one time step influences all subsequent time steps, naturally enforces temporal consistency in the transformation compared to other existing diffeomorphic registration algorithms and does not restrict the space of solutions to stationary displacement fields over all time steps. A multiresolution strategy is presented to find the optimal number of time steps and solve efficiently the joint optimization of all transformations in the registration chain. Accuracy of the algorithm in the presence of large deformations is illustrated and compared to other standard or diffeomorphic registration algorithms. Our algorithm is applied here in the context of monitoring disease progression by image-based quantification of aneurysmal growth post endovascular treatment. In this context, aneurysm volume changes occurring after embolization are quantified by integrating the Jacobian of the diffeomorphic transformation between 3D rotational angiography images acquired at subsequent followups.

Small animal registration is an important step for molecular image analysis. Skeleton registration from whole-body or only partial micro Computerized Tomography (CT) image is often performed to match individual rats to atlases and templates, for example to identify organs in positron emission tomography (PET). In this paper, we extend the shape context matching technique for 3D surface registration and apply it for rat hind limb skeleton registration from CT images. Using the proposed method, after standard affine iterative closest point (ICP) registration, correspondences between the 3D points from sour and target objects were robustly found and used to deform the limb skeleton surface with thin-plate-spline (TPS). Experiments are described using phantoms and actual rat hind limb skeletons. On animals, mean square errors were decreased by the proposed registration compared to that of its initial alignment. Visually, skeletons were successfully registered even in cases of very different animal poses.

In this work we present a novel approach for elastic image registration of multi-phase contrast enhanced CT images of liver. A problem in registration of multiphase CT is that the images contain similar but complementary structures. In our application each image shows a different part of the vessel system, e.g., portal/hepatic venous/arterial, or biliary vessels. Portal, arterial and biliary vessels run in parallel and abut on each other forming the so called portal triad, while hepatic veins run independent. Naive registration will tend to align complementary vessel. Our new approach is based on minimizing a cost function consisting of a distance measure and a regularizer. For the distance we use the recently proposed normalized gradient field measure that focuses on the alignment of edges. For the regularizer we use the linear elastic potential. The key feature of our approach is an additional penalty term using segmentations of the different vessel systems in the images to avoid overlaps of complementary structures. We successfully demonstrate our new method by real data examples.

The recent introduction of next generation spectral OCT scanners has enabled routine acquisition of high resolution, 3D cross-sectional volumetric images of the retina. 3D OCT is used in the detection and management of serious eye diseases such as glaucoma and age-related macular degeneration. For follow-up studies, image registration is a vital tool to enable more precise, quantitative comparison of disease states. This work presents a registration method based on a recently introduced extension of the 2D Scale-Invariant Feature Transform (SIFT) framework1 to 3D.2 The SIFT feature extractor locates minima and maxima in the difference of Gaussian scale space to find salient feature points. It then uses histograms of the local gradient directions around each found extremum in 3D to characterize them in a 4096 element feature vector. Matching points are found by comparing the distance between feature vectors. We apply this method to the rigid registration of optic nerve head- (ONH) and macula-centered 3D OCT scans of the same patient that have only limited overlap. Three OCT data set pairs with known deformation were used for quantitative assessment of the method's robustness and accuracy when deformations of rotation and scaling were considered. Three-dimensional registration accuracy of 2.0±3.3 voxels was observed. The accuracy was assessed as average voxel distance error in N=1572 matched locations. The registration method was applied to 12 3D OCT scans (200 x 200 x 1024 voxels) of 6 normal eyes imaged in vivo to demonstrate the clinical utility and robustness of the method in a real-world environment.

Registration is a primary step in tracking pathological changes in medical images. Point-based registration requires a set of distinct, identifiable and comparable landmark points to be extracted from images. In this work, we illustrate a method for obtaining landmarks based on changes in a topographic descriptor of a retinal image. Building on the curvature primal sketch introduced by Asada and Brady1 for describing interest points on planar curves, we extend the notion to grayscale images. We view an image as a topographic surface and propose to identify interest points on the surface using the surface curvature as a descriptor. This is illustrated by modeling retinal vessels as trenches and identifying landmarks as points where the trench behaviour changes, such as it splits or bends sharply. Based on this model, we present a method which uses the surface curvature to characterise landmark points on retinal vessels as points of high dispersion in the curvature orientation histogram computed around the points. This approach yields junction/crossover points of retinal vessels and provides a means to derive additional information about the type of junction. A scheme is developed for using such information and determining the correspondence between sets of landmarks from two images related by a rigid transformation. In this paper we present the details of the proposed approach and results of testing on images from public domain datasets. Results include comparison of landmark detection with two other methods, and results of correspondence derivation. Results show the approach to be successful and fast.

In this paper we present a learning-based guidewire localization algorithm which can be constrained by user inputs. The proposed algorithm automatically localizes guidewires in fluoroscopic images. In cases where the results are not satisfactory, the user can provide input to constrain the algorithm by clicking on the guidewire segment missed by the detection algorithm. The algorithm then re-localizes the guidewire and updates the result in less than 0.3 second. In extreme cases, more constraints can be provided until a satisfactory result is reached. The proposed algorithm can not only serve as an efficient initialization tool for guidewire tracking, it can also serve as an efficient annotation tool, either for cardiologists to mark the guidewire, or to build up a labeled database for evaluation. Through the improvement of the initialization of guidewire tracking, it also helps to improve the visibility of the guidewire during interventional procedures. Our study shows that even highly complicated guidewires can mostly be localized within 5 seconds by less than 6 clicks.

In this paper, we present a novel hierarchical framework of guidewire tracking for image-guided interventions. Our method can automatically and robustly track a guidewire in fluoroscopy sequences during interventional procedures. The method consists of three main components: learning based guidewire segment detection, robust and fast rigid tracking, and nonrigid guidewire tracking. Each component aims to handle guidewire motion at a specific level. The learning based segment detection identifies small segments of a guidewire at the level of individual frames, and provides unique primitive features for subsequent tracking. Based on identified guidewire segments, the rigid tracking method robustly tracks the guidewire across successive frames, assuming that a major motion of guidewire is rigid, mainly caused by the breathing motion and table movement. Finally, a non-rigid tracking algorithm is applied to finely deform the guidewire to provide accurate shape. The presented guidewire tracking method has been evaluated on a test set of 47 sequences with more than 1000 frames. Quantitative evaluation demonstrates that the mean tracking error on the guidewire body is less than 2 pixels. Therefore the presented guidewire tracking method has a great potential for applications in image guided interventions.

During interventional radiology procedures, guide-wires are usually inserted into the patients vascular tree for diagnosis or healing purpose. These procedures are monitored with an Xray interventional system providing images of the interventional devices navigating through the patient's body. The automatic detection of such tools by image processing means has gained maturity over the past years and enables applications ranging from image enhancement to multimodal image fusion. Sophisticated detection methods are emerging, which rely on a variety of device enhancement techniques. In this article we reviewed and classified these techniques into three families. We chose a state of the art approach in each of them and built a rigorous framework to compare their detection capability and their computational complexity. Through simulations and the intensive use of ROC curves we demonstrated that the Hessian based methods are the most robust to strong curvature of the devices and that the family of rotated filters technique is the most suited for detecting low CNR and low curvature devices. The steerable filter approach demonstrated less interesting detection capabilities and appears to be the most expensive one to compute. Finally we demonstrated the interest of automatic guide-wire detection on a clinical topic: the compensation of respiratory motion in multimodal image fusion.

Tumor segmentation from MRI data is a particularly challenging and time consuming task. Tumors have a large diversity in shape and appearance with intensities overlapping the normal brain tissues. In addition, an expanding tumor can also deflect and deform nearby tissue. Our work addresses these last two difficult problems. We use the available MRI modalities (T1, T1c, T2) and their texture characteristics to construct a multi-dimensional feature set. Further, we extract clusters which provide a compact representation of the essential information in these features. The main idea in this paper is to incorporate these clustered features into the 3D variational segmentation framework. In contrast to the previous variational approaches, we propose a segmentation method that evolves the contour in a supervised fashion. The segmentation boundary is driven by the learned inside and outside region voxel probabilities in the cluster space. We incorporate prior knowledge about the normal brain tissue appearance, during the estimation of these region statistics. In particular, we use a Dirichlet prior that discourages the clusters in the ventricles to be in the tumor and hence better disambiguate the tumor from brain tissue. We show the performance of our method on real MRI scans. The experimental dataset includes MRI scans, from patients with difficult instances, with tumors that are inhomogeneous in appearance, small in size and in proximity to the major structures in the brain. Our method shows good results on these test cases.

We present a novel framework for the simultaneous segmentation of multiple interacting surfaces belonging to multiple mutually interacting objects. The method is a non-trivial extension of our previously reported optimal multi-surface segmentation. Considering an example application of knee-cartilage segmentation, the framework consists of the following main steps: 1) Shape model construction: Building a mean shape for each bone of the joint (femur, tibia, patella) from interactively segmented volumetric datasets. Using the resulting mean-shape model - identification of cartilage, non-cartilage, and transition areas on the mean-shape bone model surfaces. 2) Presegmentation: Employment of iterative optimal surface detection method to achieve approximate segmentation of individual bone surfaces. 3) Cross-object surface mapping: Detection of inter-bone equidistant separating sheets to help identify corresponding vertex pairs for all interacting surfaces. 4) Multi-object, multi-surface graph construction and final segmentation: Construction of a single multi-bone, multi-surface graph so that two surfaces (bone and cartilage) with zero and non-zero intervening distances can be detected for each bone of the joint, according to whether or not cartilage can be locally absent or present on the bone. To define inter-object relationships, corresponding vertex pairs identified using the separating sheets were interlinked in the graph. The graph optimization algorithm acted on the entire multiobject, multi-surface graph to yield a globally optimal solution. The segmentation framework was tested on 16 MR-DESS knee-joint datasets from the Osteoarthritis Initiative database. The average signed surface positioning error for the 6 detected surfaces ranged from 0.00 to 0.12 mm. When independently initialized, the signed reproducibility error of bone and cartilage segmentation ranged from 0.00 to 0.26 mm. The results showed that this framework provides robust, accurate, and reproducible segmentation of the knee joint bone and cartilage surfaces of the femur, tibia, and patella. As a general segmentation tool, the developed framework can be applied to a broad range of multi-object segmentation problems.

We introduce a new approach for tracking-based segmentation of 3D tubular structures. The approach is based on a novel combination of a 3D cylindrical intensity model and particle filter tracking. In comparison to earlier work we utilize a 3D intensity model as the measurement model of the particle filter, thus a more realistic 3D appearance model is used that directly represents the image intensities of 3D tubular structures within semiglobal regions-of-interest. We have successfully applied our approach using 3D synthetic images and real 3D MRA image data of the human pelvis.

In the last years, bioimaging has turned from qualitative measurements towards a high-throughput and highcontent modality, providing multiple variables for each biological sample analyzed. We present a system which combines machine learning based semantic image annotation and visual data mining to analyze such new multivariate bioimage data. Machine learning is employed for automatic semantic annotation of regions of interest. The annotation is the prerequisite for a biological object-oriented exploration of the feature space derived from the image variables. With the aid of visual data mining, the obtained data can be explored simultaneously in the image as well as in the feature domain. Especially when little is known of the underlying data, for example in the case of exploring the effects of a drug treatment, visual data mining can greatly aid the process of data evaluation. We demonstrate how our system is used for image evaluation to obtain information relevant to diabetes study and screening of new anti-diabetes treatments. Cells of the Islet of Langerhans and whole pancreas in pancreas tissue samples are annotated and object specific molecular features are extracted from aligned multichannel fluorescence images. These are interactively evaluated for cell type classification in order to determine the cell number and mass. Only few parameters need to be specified which makes it usable also for non computer experts and allows for high-throughput analysis.

American College of Radiology introduces a standard in classification, the breast imaging reporting and data system (BIRADS), standardize the reporting of ultrasound findings, clarify its interpretation, and facilitate communication between clinicians. The effective use of new technologies to support healthcare initiatives is important and current research is moving towards implementing computer tools in the diagnostics process. Initially a detailed study was carried out to evaluate the performance of two commonly used appearance based classification algorithms, based on the use of Principal Component Analysis (PCA), and two dimensional linear discriminant analysis (2D-LDA). The study showed that these two appearance based classification approaches are not capable of handling the classification of ultrasound breast image lesions. Therefore further investigations in the use of a popular feature based classifier - Support Vector Machine (SVM) was conducted. A pre-processing step before feature based classification is feature extraction, which involve shape, texture and edge descriptors for the Region of Interest (ROI). The input dataset to SVM classification is from a fully automated ROI detection. We achieve the success rate of 0.550 in PCA, 0.500 in LDA, and 0.931 in SVM. The best combination of features in SVM classification is to combine the shape, texture and edge descriptors, with sensitivity 0.840 and specificity 0.968. This paper briefly reviews the background to the project and then details the ongoing research. In conclusion, we discuss the contributions, limitations, and future plans of our work.

A glioma is a type of cancer occurring, in the majority of cases, in the brain. The World Health Organization (WHO) assigns a grade from I to IV to this tumor, with I being the least aggressive and IV being the most aggressive. In glioma cells of grade IV the Epidermal Growth Factor Receptors (EGFRs) are over expressed. In this paper we hypothesize that this overexpression occurs also for gliomas of grades I to III. Moreover, we present a medical study aiming to determine the correlation between the WHO classification and the EGFR quantity in glioma tissue. We define five quantity classes for EGFR. First, results of immunohistochemical staining on brain glioma slices, which visualize the EGFR quantity, are examined under an optical microscope and manually classified into these five classes. In this paper we propose to perform this classification automatically using statistical pattern recognition technique for digital images. For this, digital microscope images of glioma are acquired and their histograms computed. Afterwards, all five EGFR quantity classes (image classes) are statistically modeled using training samples. This allows a fully automatic classification of unknown images into one of the five classes using the Maximum Likelihood (ML) estimation. Experimental results show that, on the one hand, the automatic EGFR quantity classification performs with a quite high accuracy, on the other hand, it is done much faster than manual labeling done by a human.

Feature extraction and selection are of great importance in neuroimage classification for identifying informative features and reducing feature dimensionality, which are generally implemented as two separate steps. This paper presents an integrated feature extraction and selection algorithm with two iterative steps: constrained subspace learning based feature extraction and support vector machine (SVM) based feature selection. The subspace learning based feature extraction focuses on the brain regions with higher possibility of being affected by the disease under study, while the possibility of brain regions being affected by disease is estimated by the SVM based feature selection, in conjunction with SVM classification. This algorithm can not only take into account the inter-correlation among different brain regions, but also overcome the limitation of traditional subspace learning based feature extraction methods. To achieve robust performance and optimal selection of parameters involved in feature extraction, selection, and classification, a bootstrapping strategy is used to generate multiple versions of training and testing sets for parameter optimization, according to the classification performance measured by the area under the ROC (receiver operating characteristic) curve. The integrated feature extraction and selection method is applied to a structural MR image based Alzheimer's disease (AD) study with 98 non-demented and 100 demented subjects. Cross-validation results indicate that the proposed algorithm can improve performance of the traditional subspace learning based classification.

Being able to automatically determine which portion of the human body is shown by a CT volume image offers various possibilities like automatic labeling of images or initializing subsequent image analysis algorithms. This paper presents a method that takes a CT volume as input and outputs the vertical body coordinates of its top and bottom slice in a normalized coordinate system whose origin and unit length are determined by anatomical landmarks. Each slice of a volume is described by a histogram of visual words: Feature vectors consisting of an intensity histogram and a SURF descriptor are first computed on a regular grid and then classified into the closest visual words to form a histogram. The vocabulary of visual words is a quantization of the feature space by offline clustering a large number of feature vectors from prototype volumes into visual words (or cluster centers) via the K-Means algorithm. For a set of prototype volumes whose body coordinates are known the slice descriptions are computed in advance. The body coordinates of a test volume are computed by a 1D rigid registration of the test volume with the prototype volumes in axial direction. The similarity of two slices is measured by comparing their histograms of visual words. Cross validation on a dataset of 44 volumes proved the robustness of the results. Even for test volumes of ca. 20cm height, the average error was 15.8mm.

One major hallmark of the Alzheimer's disease (AD) is the loss of neurons in the brain. In many cases, medical experts use magnetic resonance imaging (MRI) to qualitatively measure the neuronal loss by the shrinkage or enlargement of the structures-of-interest. Brain ventricle is one of the popular choices. It is easily detectable in clinical MR images due to the high contrast of the cerebro-spinal fluid (CSF) with the rest of the parenchyma. Moreover, atrophy in any periventricular structure will directly lead to ventricle enlargement. For quantitative analysis, volume is the common choice. However, volume is a gross measure and it cannot capture the entire complexity of the anatomical shape. Since most existing shape descriptors are complex and difficult-to-reproduce, more straightforward and robust ways to extract ventricle shape features are preferred in the diagnosis. In this paper, a novel ventricle shape based classification method for Alzheimer's disease has been proposed. Training process is carried out to generate two probability maps for two training classes: healthy controls (HC) and AD patients. By subtracting the HC probability map from the AD probability map, we get a 3D ventricle discriminant map. Then a matching coefficient has been calculated between each training subject and the discriminant map. An adjustable cut-off point of the matching coefficients has been drawn for the two classes. Generally, the higher the cut-off point that has been drawn, the higher specificity can be achieved. However, it will result in relatively lower sensitivity and vice versa. The benchmarked results against volume based classification show that the area under the ROC curves for our proposed method is as high as 0.86 compared with only 0.71 for volume based classification method.

Independent Component Analysis (ICA) is a blind source separation technique that has previously been applied to various time-varying signals. It may in particular be utilized to study 1H-MR spectroscopic imaging (MRSI) data. The work presented firstly investigates preprocessing and parameterization for ICA on simulated data to assess different strategies. We then applied ICA processing to 2D/3D brain and prostate MRSI data obtained from two healthy volunteers and 17 patients. We conducted a correlation analysis of the mixing and separating matrices resulting from ICA processing with maps obtained from metabolite quantitations in order to elucidate the relationship between quantitative and ICA results. We found that the mixing matrices corresponding to the estimated independent components highly correlate with the metabolite maps for some cases, and for others differ. We provide explanations and speculations for that and propose a scheme to utilize the knowledge for hot-spot detection. From our experience, ICA is much faster than the calculation of metabolic maps. Additionally, water and lipid contaminations are on the way removed from the data; the user needs not manually exclude spectroscopic voxels from processing or analysis. ICA results show hot spots in the data, even where quantitation-based metabolic maps are difficult to assess due to noisy data or macromolecule distortions.

Diffusion imaging provides the ability to study white matter connectivity and integrity noninvasively. Diffusion weighted imaging contains orientation information that must be appropriately reoriented when applying spatial transforms to the resulting imaging data. Alexander et al. have introduced two methods to resolve the reorientation problem. In the first method, the rotation matrix is computed from the transform and the tensors are reoriented. The second method called as the preservation of principal direction (PPD) method, takes into account the deformation and rotation components to estimate the rotation matrix. These methods cannot be directly used for higher order diffusion models (e.g. Q-ball). We have introduced a novel technique called gradient rotation where the rotation is directly applied to the diffusion sensitizing gradients providing a voxel by voxel estimate of the diffusion gradients instead of a volume of by volume estimate. A PPD equivalent gradient rotation can be computed using principal component analysis (PCA). Four subjects were spatially normalized to a template subject using a multistage registration sequence that includes nonlinear diffeomorphic demons registration. Comparative results of all four methods have been shown. It can be observed that all the methods work closely to each other, PPD (original and gradient equivalent) being slightly better than rigid rotation, based on the fact that it includes the shear and scale component. Results also demonstrate that the multistage registration is a viable method for spatial normalization of diffusion models.

Diffusion MRI has become an established research tool for the investigation of tissue structure and orientation from which has stemmed a number of variations, such as Diffusion Tensor Imaging (DTI) Diffusion Spectrum Imaging (DSI) and QBall Imaging (QBI). The acquisition and analysis of such data is very challenging due to its complexity. Recently, an exciting new Kalman filtering framework has been proposed for DTI and QBI reconstructions in real time during the repetition time (TR) of the acquisition sequence. In this article, we first revisit and thoroughly analyze this approach and show it is actually sub-optimal and not recursively minimizing the intended criterion due to the Laplace-Beltrami regularization term. Then, we propose a new approach that implements the QBI reconstruction algorithm in real-time using a fast and robust Laplace-Beltrami regularization without sacrificing the optimality of the Kalman filter. We demonstrate that our method solves the correct minimization problem at each iteration and recursively provides the optimal QBI solution. We validate with real QBI data that our proposed real-time method is equivalent in terms of QBI estimation accuracy to the standard off-line processing techniques and outperforms the existing solution. This opens new and interesting opportunities for real-time feedback for clinicians during an acquisition and also for researchers investigating into optimal diffusion orientation sets and, real-time fiber tracking and connectivity mapping.

Diffusion weighted imaging (DWI) technique has been used to help understand the human brain white matter fiber structures in vivo. Currently used standard diffusion tensor magnetic resonance imaging (DTI) tractography based on the second order diffusion tensor model has limitations in its ability to resolve complex fiber tracts. The generalized diffusion tensor (GDT) imaging technique has been proposed to overcome these limitations associated with the standard second order tensor model. Based on the GDT model, a generalized partial differential equation (PDE) governing the anisotropic diffusion process can be derived. For the purpose of solving the PDE and computing the generalized diffusion tensor, we derive a generalized analytic expression for the high order b matrix in the case of twice-refocused spin echo (TRSE) pulse sequence which is used in the DWI data acquisition. The TRSE pulse sequence is considered because of its ability to null the eddy current effect generated during the scanning. The b matrix was computed by integrating the transverse precessing magnetization between the excitation time and the echo time (TE). In our experiments, we show some computational results of the generalized b matrix based on the new analytic expression. In addition, comparisons between the generalized b matrix computed using our formula and the second order b matrix given by the MRI machine are presented. The characteristics of the fomula and the data are discussed at last.

This work introduces a MATLAB-based tool developed for investigating functional connectivity in the brain. Independent component analysis (ICA) is used as a measure of voxel similarity which allows the user to find and view statistically independent maps of correlated voxels. These maps of correlated voxel activity may indicate functionally connected regions. Specialized clustering and feature extraction techniques have been designed to find and characterize clusters of activated voxels, which allows comparison of the spatial maps of correlation across subjects. This method is also used to compare the ICA generated images to fMRI images showing statistically significant activations generated by Statistical Parametric Mapping (SPM). The capability of querying specific coordinates in the brain supports integration and comparison with other data modalities such as Cortical Stimulation Mapping and Single Unit Recordings.

Ultrasound image quality is degraded because of the presence of speckle, which causes loss of image contrast resolution and makes the detection of small features difficult. The traditional nonlinear diffusion filtering of speckle reduction with explicit schemes can achieve desirable results, but they are only stable for very small time steps. Semi-implicit additive operator splitting (AOS) schemes for nonlinear diffusion are stable for all time size and more efficient than the traditional explicit schemes. However, the AOS schemes are still inefficient for real time speckle reducing of ultrasound images. Current graphics processing units (GPUs) offers an opportunity to boost the computation speed of AOS schemes through high computational power at low cost. In this paper, an optimized GPU framework for AOS schemes is presented. By using the well-established method of cyclic reduction of tridiagonal systems in our framework, we are able to implement the AOS schemes on GPU. Experiments from CPU implemented AOS schemes and our GPU based framework show that our method is about 10 times faster than the CPU implementation. Our presented framework deals with the local coherence anisotropic diffusion, but it can be generalized to the class of nonlinear diffusion methods which can be discretized by AOS schemes.

Computer-aided detection (CAD) is a computerized procedure in medical science that supports the medical team's interpretations and decisions. CAD often uses information from a medical imaging modality such as Computed Tomography to detect suspicious lesions. Algorithms to detect these lesions are based on geometric models which can describe the local structures and thus provide potential region candidates. Geometrical descriptive models are very dependent on the data quality which may affect the false positive rates in CAD. In this paper we propose an efficient adaptive diffusion technique that adaptively controls the diffusion flux of the local structures in the data using robust statistics. The proposed method acts isotropically in the homogeneous regions and anisotropically in the vicinity of jump discontinuities. This method structurally enhances the data and makes the geometrical descriptive models robust. For the iterative solver, we use an efficient gradient descent flow solver based on a PDE formulation of the problem. The whole proposed strategy, which makes use of adaptive diffusion filter coupled with gradient descent flows has been developed and evaluated on clinical data in the application to colonic polyp detection in Computed Tomography Colonography.

Implantable cardiac devices, such as stents and septal defect closure devices are sometimes difficult to see on angiographic X-ray projection images. We present a method to enhance the visibility of these devices in rotational X-ray angiography acquisitions using automated marker detection and motion compensation. Automatic marker detection allows registration of the devices in the images of the rotational run. Motion compensation is done by warping the images to a specific reference position. Averaging multiple of those motion compensated images together with the reference frame results in an enhanced image with improved visibility due to an increase in contrast of the device with the background structure. This allows the clinician to look at the device from multiple angles with an improved visibility of the device to better appreciate the 3D geometry of the device. In particular, enhancement of rotational acquisitions compared to standard enhanced fixed-angle acquisitions allows the clinician to better perceive any asymmetry in the deployed device.

Clinical MRI images are generally corrupted by random noise during acquisition with blurred subtle structure features. Many denoising methods have been proposed to remove noise from corrupted images at the expense of distorted structure features. Therefore, there is always compromise between removing noise and preserving structure information for denoising methods. For a specific denoising method, it is crucial to tune it so that the best tradeoff can be obtained. In this paper, we define several cost functions to assess the quality of noise removal and that of structure information preserved in the denoised image. Strength Pareto Evolutionary Algorithm 2 (SPEA2) is utilized to simultaneously optimize the cost functions by modifying parameters associated with the denoising methods. The effectiveness of the algorithm is demonstrated by applying the proposed optimization procedure to enhance the image denoising results using block matching and 3D collaborative filtering. Experimental results show that the proposed optimization algorithm can significantly improve the performance of image denoising methods in terms of noise removal and structure information preservation.

Digital mammography images often contain noise proportional to the dose given to the patient. The signal to noise ratio is typically improved by increasing the dose as noise reduction methods tend to blur clinically important edges and details, such as fibers or calcifications. A new algorithm is presented to reduce noise while preserving features relevant in digital mammography and allowing for an overall reduction of patient dose.

Visualization of subcutaneous veins is very difficult with the naked eye, but important for diagnosis of medical conditions and different medical procedures such as catheter insertion and blood withdrawal. Moreover, recent studies showed that the images of subcutaneous veins could be used for biometric identification. The majority of methods used for enhancing the contrast between the subcutaneous veins and surrounding tissue are based on simple imaging systems utilizing CMOS or CCD cameras with LED illumination capable of acquiring images from the near infrared spectral region, usually near 900 nm. However, such simplified imaging methods cannot exploit the full potential of the spectral information. In this paper, a new highly versatile method for enhancing the contrast of subcutaneous veins based on state-of-the-art high-resolution hyper-spectral imaging system utilizing the spectral region from 550 to 1700 nm is presented. First, a detailed analysis of the contrast between the subcutaneous veins and the surrounding tissue as a function of wavelength, for several different positions on the human arm, was performed in order to extract the spectral regions with the highest contrast. The highest contrast images were acquired at 1100 nm, however, combining the individual images from the extracted spectral regions by the proposed contrast enhancement method resulted in a single image with up to ten-fold better contrast. Therefore, the proposed method has proved to be a useful tool for visualization of subcutaneous veins.

Accurate quantification of positron emission tomography (PET) is important for diagnosis and assessment of cancer treatment. The low spatial resolution of PET imaging induces partial volume effect to PET images that biases quantification. A PET partial volume correction method is proposed using high-resolution, anatomical information from magnetic resonance images (MRI). The corrected PET is pursued by removing the convolution of PET point spread function (PSF) and by preserving edges present in PET and the aligned MR images. The correction is implemented in a Bayesian's deconvolution framework that is minimized by a conjugate gradient method. The method is evaluated on simulated phantom and brain PET images. The results show that the method effectively restores 102 ± 7% of the true PET activity with a size of greater than the full-width at half maximum of the point spread function. We also applied the method to synthesized brain PET data. The method does not require prior information about tracer activity within tissue regions. It can offer a partial volume correction method for various PET applications and can be particularly useful for combined PET/MRI studies.

Flat-panel detector-based cone beam CT usually employs FDK algorithm as the reconstruction method. Traditionally, the row-wise ramp linear filtering was regularized by noise-suppression windows, such as Shepp-Logan, Hamming windows etc before the backprojection to get the final acceptable (in terms of SNR) reconstructed 3-D volume data. Though noise was reduced, this linear filtering regularized by noise suppression window had the potential to affect the signal spatial resolution and thus to reduce the sharpness of the structure boundaries within the breast image especially impeding the detection of the small calcifications and very small abnormalities that may indicate early breast cancer. Furthermore, the reconstructed images were still characterized by smudges. In order to combat the aforementioned shortcomings, a Wavelet regularization method was conducted on projection data followed by row-wise ramp linear filtering inherited within FDK.

The purpose of this study is to develop an automated method without user interaction for the optimal placing of ROI for selecting optimal data acquisition window in coronary MRA. One of the major problems of magnetic resonance coronary angiography (MRCA) is the effective suppression of coronary motion due to respiration and cardiac contraction. To compensate for cardiac movement, data acquisition is generally limited to the coronary artery rest period mainly found during end-diastole. End-diastole is called cardiac rest period. Generally, cardiac rest period is decided by operator. Therefore it is subjective, and requires many experiences. As for setting region of interest right coronary is known as an appropriate region for determining cardiac rest period. We proposed a method that is based on the extraction of intensive change regions and the calculation of the correlation coefficient. We tested the algorithm for two sets of clinical MR images and the results are shown.

3D coronary modeling extracts the centerlines and width of the coronary arteries from a rotational sequence of angiographies. This process heavily relies on a preliminary filtering of the 2D angiograms that enhances the vessels. We propose an improved vessel enhancement method specifically designed for this application. It keeps the advantages of Hessian-based extraction methods (speed, robustness, multiscale) while bypassing its more important limitations: the blurring of bifurcations, and the incomplete filling of very large vessels. The major contributions of this paper are twofold. First, the classical centered kernel used in Hessian-based methods is substituted with an elongated off-centered kernel. The new filter detects the different orientations involved at a bifurcation: it can answer properly to 'half vessels' beginning at the considered pixel (as opposed to the centered classical filter). The proposed "semi-oriented ridge" filter is also more robust to noise, and it stays multi-scale and quickly computable. Second, an original bifurcation detection and enhancement method is presented, based on the following heuristics: "bifurcations have three vessels (at least) in their immediate neighborhood". More precisely, the semi-oriented ridges answers in each tested orientation θ∈]-π,π] are stored in a circular histogram. The proposed bifurcation energy is the height of the third peak in this histogram: it will have a significant value at bifurcations only. The performance of the complete framework is demonstrated both on the produced vessel maps and on the final modeling results.

Tridirectionally encoded phase-contrast MRI is a technique to non-invasively acquire time-resolved velocity vector fields of blood flow. These may not only be used to analyze pathological flow patterns, but also to quantify flow at arbitrary positions within the acquired volume. In this paper we examine the validity of this approach by analyzing the consistency of related quantifications instead of comparing it with an external reference measurement. Datasets of the thoracic aorta were acquired from 6 pigs, 1 healthy volunteer and 3 patients with artificial aortic valves. Using in-house software an elliptical flow quantification plane was placed manually at 6 positions along the descending aorta where it was rotated to 5 different angles. For each configuration flow was computed based on the original data and data that had been corrected for phase offsets. Results reveal that quantifications are more dependent on changes in position than on changes in angle. Phase offset correction considerably reduces this dependency. Overall consistency is good with a maximum variation coefficient of 9.9% and a mean variation coefficient of 7.2%.

Unsupervised analysis of time-series of live-cell images is one of the important tools of quantitative biology. Due to permanent cell motility or displacements of subcellular structures, microscopic images exhibit intrinsic non-uniform motion. In this article, we present a novel approach for detection of non-uniform multi-body motion which is based on combination of the Fourier-phase correlation with iterative probing target and background image regions similar to the strategy known from saccadic eye movements. We derive theoretical expressions that yield plausible explanation why this strategy turns out to be advantageous for tracking particular image pattern. Our experiments with synthetic and live-cell images demonstrate that the proposed approach is capable of accurately detecting non-uniform motion in synthetic and live-cell images.

We present a post-reconstruction motion-compensated spatio-temporal filtering method for noise reduction in cardiac gated SPECT images. SPECT imaging suffers from low photon count due to radioactive dose limitations resulting in a high noise level in the reconstructed images. This is especially true in gated cardiac SPECT where the total number of counts is divided into a number of gates (time frames). Classical spatio-temporal filtering approaches, used in gated cardiac SPECT for noise reduction, do not accurately account for myocardium motion and brightening and therefore perform sub-optimally. The proposed post-reconstruction method consists of two steps: motion and brightening estimation and spatio-temporal motion-compensated filtering. In the first step we utilize a left ventricle model and a deformable mesh structure. The second step, which consists of motion-compensated spatio-temporal filtering, makes use of estimated myocardial motion to enable accurate smoothing. Additionally, the algorithm preserves myocardial brightening, a result of partial volume effect which is widely used as a diagnostic feature. The proposed method is evaluated quantitatively to assess noise reduction and the influence on estimated ejection fraction.

Recently Strain and strain rate imaging have proved their superiority with respect to classical motion estimation methods in myocardial evaluation as a novel technique for quantitative analysis of myocardial function. Here in this paper, we propose a novel strain rate imaging algorithm using a new optical flow technique which is more rapid and accurate than the previous correlation-based methods. The new method presumes a spatiotemporal constancy of intensity and Magnitude of the image. Moreover the method makes use of the spline moment in a multiresolution approach. Moreover cardiac central point is obtained using a combination of center of mass and endocardial tracking. It is proved that the proposed method helps overcome the intensity variations of ultrasound texture while preserving the ability of motion estimation technique for different motions and orientations. Evaluation is performed on simulated, phantom (a contractile rubber balloon) and real sequences and proves that this technique is more accurate and faster than the previous methods.

We propose a cartilage matching technique based on the registration of the corresponding bone structures instead of using the cartilage. Our method consists of five steps. First, cartilage and corresponding bone structures are extracted by semi-automatic segmentation. Second, gross translational mismatch between corresponding bone structures is corrected by point-based rough registration. The center of inertia (COI) of each segmented bone structure is considered as the reference point. Third, the initial alignment is refined by distance-based surface registration. For fast and robust convergence of the distance measure to the optimal value, a 3D distance map is generated by the Gaussian-weighted narrow-band distance propagation. Fourth, rigid transformation of the bone surface registration is applied to the cartilage of baseline MR images. Finally, morphological differences of the corresponding cartilages are visualized by color-coded mapping and image fusion. Experimental results show that the cartilage morphological changes of baseline and follow-up MR knee images can be easily recognized by the correct registration of the corresponding bones.

This paper evaluates the performance of two non-rigid image registration techniques. The moving least squares (MLS) technique is compared to the more common thin-plate spline (TPS) method. Both methods interpolate a set of fiducial points in registering two images. An attractive feature of the MLS method is that it seeks to minimize local scaling and shearing, producing a global transformation that is as rigid as possible. The MLS and TPS techniques are applied to twoand three-dimensional medical images. Both qualitative and quantitative comparisons are presented. The two techniques are quantitatively evaluated by computing target registration errors (TREs) at selected points of interest. Our results indicate that the MLS algorithm performs better than the TPS method with lower TRE values and visually better registered images, indicating that MLS may be a better candidate for registration tasks when rigid registration is insufficient but the deformation field is sought to be minimal.

Integration and correlation of brain's electrical (EEG) and physiological activity (PET, fMRI) is crucial for the early evaluation of patients with neurophysiological disorders, such as epilepsy. Based on the scalp-recorded EEG signals, the source image of brain's electrical activity can be reconstructed and spatially correlated with tomographic functional images, thereby aiding to the characterization and localization of epileptic foci. However, mis-localization of the electrode positions, with respect to the underlying anatomy, adversely affects the localization precision performed by the interpretation of the source image. In this paper, a novel method for registration of EEG electrode positions to tomographic functional images of the brain is proposed. Accuracy and robustness of the registration were evaluated on three databases of real and simulated PET and real fMRI images. The registration method showed good convergence properties for both PET [>10 mm] and especially fMRI images [>30 mm]. Based on Monte Carlo simulations, the obtained mean registration error of electrode positions in tomographic functional images was in the range of 1-2 corresponding voxel size. In this way, the constant bias in the reconstructed source image, that is due to the mis-registration of EEG electrode positions, can be suppressed with respect to the random errors induced by EEG signal noise. Finally, we aim to improve, or at all enable, the integration and application of the many functional modalities involved in the analysis and evaluation of clinical neurophysiological disorders.

Coregistration of in vivo magnetic resonance imaging (MRI) with histology provides validation of disease biomarker and pathobiology studies. Although thin-plate splines are widely used in such image registration, point landmark selection is error prone and often time-consuming. We present a technique to optimize landmark selection for thin-plate splines and demonstrate its usefulness in warping rodent brain MRI to histological sections. In this technique, contours are drawn on the corresponding MRI slices and images of histological sections. The landmarks are extracted from the contours by equal spacing then optimized by minimizing a cost function consisting of the landmark displacement and contour curvature. The technique was validated using simulation data and brain MRI-histology coregistration in a murine model of HIV-1 encephalitis. Registration error was quantified by calculating target registration error (TRE). The TRE of approximately 8 pixels for 20-80 landmarks without optimization was stable at different landmark numbers. The optimized results were more accurate at low landmark numbers (TRE of approximately 2 pixels for 50 landmarks), while the accuracy decreased (TRE approximately 8 pixels for larger numbers of landmarks (70- 80). The results demonstrated that registration accuracy decreases with the increasing landmark numbers offering more confidence in MRI-histology registration using thin-plate splines.

Virtual microscopy has the potential to partially replace traditional microscopy. For virtualization, the slide is scanned once by a fully automatized robotic microscope and saved digitally. Typically, such a scan results in several hundreds to thousands of fields of view. Since robotic stages have positioning errors, these fields of view have to be registered locally and globally in an additional step. In this work we propose a new global mosaicking method for the creation of virtual slides based on sub-pixel exact phase correlation for local alignment in combination with Prim's minimum spanning tree algorithm for global alignment. Our algorithm allows for a robust reproduction of the original slide even in the presence of views with little to no information content. This makes it especially suitable for the mosaicking of cervical smears. These smears often exhibit large empty areas, which do not contain enough information for common stitching approaches.

We present a method to automate acquisition of MR brain scans to allow consistent alignment of diagnostic images for patient follow-up, and to depict standardized anatomy for all patients. The algorithm takes as input a low-resolution acquisition that depicts the patient position within the scanner. The mid-sagittal plane dividing the brain hemispheres is automatically detected, as are bony landmarks at the front and back of the skull. The orientation and position of a subsequent diagnostic, high resolution scan is then aligned based on these landmarks. The method was tested on 91 data sets, and was completely successful in 93.4% of cases, performed acceptably in 4.4% of cases, and failed for 1.1%. We conclude that the method is suitable for clinical use and should prove valuable for improving consistency of acquisitions.

We present a new method for fully automatic non-rigid registration of multimodal imagery, including structural and functional data, that utilizes multiple texutral feature images to drive an automated spline based non-linear image registration procedure. Multimodal image registration is significantly more complicated than registration of images from the same modality or protocol on account of difficulty in quantifying similarity between different structural and functional information, and also due to possible physical deformations resulting from the data acquisition process. The COFEMI technique for feature ensemble selection and combination has been previously demonstrated to improve rigid registration performance over intensity-based MI for images of dissimilar modalities with visible intensity artifacts. Hence, we present here the natural extension of feature ensembles for driving automated non-rigid image registration in our new technique termed Collection of Image-derived Non-linear Attributes for Registration Using Splines (COLLINARUS). Qualitative and quantitative evaluation of the COLLINARUS scheme is performed on several sets of real multimodal prostate images and synthetic multiprotocol brain images. Multimodal (histology and MRI) prostate image registration is performed for 6 clinical data sets comprising a total of 21 groups of in vivo structural (T2-w) MRI, functional dynamic contrast enhanced (DCE) MRI, and ex vivo WMH images with cancer present. Our method determines a non-linear transformation to align WMH with the high resolution in vivo T2-w MRI, followed by mapping of the histopathologic cancer extent onto the T2-w MRI. The cancer extent is then mapped from T2-w MRI onto DCE-MRI using the combined non-rigid and affine transformations determined by the registration. Evaluation of prostate registration is performed by comparison with the 3 time point (3TP) representation of functional DCE data, which provides an independent estimate of cancer extent. The set of synthetic multiprotocol images, acquired from the BrainWeb Simulated Brain Database, comprises 11 pairs of T1-w and proton density (PD) MRI of the brain. Following the application of a known warping to misalign the images, non-rigid registration was then performed to recover the original, correct alignment of each image pair. Quantitative evaluation of brain registration was performed by direct comparison of (1) the recovered deformation field to the applied field and (2) the original undeformed and recovered PD MRI. For each of the data sets, COLLINARUS is compared with the MI-driven counterpart of the B-spline technique. In each of the quantitative experiments, registration accuracy was found to be significantly (p < 0.05) for COLLINARUS compared with MI-driven B-spline registration. Over 11 slices, the mean absolute error in the deformation field recovered by COLLINARUS was found to be 0.8830 mm.

The task of registering 3D medical images is very computationally expensive. With CPU-based implementations of registration algorithms it is typical to use various approximations, such as subsampling, to maintain reasonable computation times. This may however result in suboptimal alignments. With the constant increase of capabilities and performances of GPUs (Graphics Processing Unit), these highly vectorized processors have become a viable alternative to CPUs for image related computation tasks. This paper describes new strategies to implement on GPU the computation of image similarity metrics for intensity-based registration, using in particular the latest features of NVIDIA's GeForce 8 architecture and the Cg language. Our experimental results show that the computations are many times faster. In this paper, several GPU implementations of two image similarity criteria for both intramodal and multi-modal registration have been compared. In particular, we propose a new efficient and flexible solution based on the geometry shader.

This paper presents a novel method to reduce the effects of interleaving motion artefacts in single-plane MR scanning of the pelvic region without the need for k-space information. Interleaved image (or multipacket) acquisition is frequently used to reduce cross-talk and scanning time during full pelvic MR scans. Patient motion during interleaved acquisition can result in non-linear "staircase" imaging artefacts which are most visible on sagittal and coronal reconstructions. These artefacts can affect the segmentation of organs, registration, and visualization. A fast method has been implemented to replace artefact affected slices in a packet with interpolated slices based on Penney et al (2004) whose method involves the registration of neighbouring slices to obtain correspondences, followed by linear interpolation of voxel intensities along the displacement fields. This interpolation method has been applied to correct motion affected MRI volumes by firstly creating a new volume where every axial slice from the artefact affected packet is removed and replaced with an interpolated slice and then secondly for each of these slices, 2D non-rigid registration is used to register each original axial slice back to its matching interpolated slice. Results show visible improvements in artefacts particularly in sagittal and coronal image reconstructions, and should result in improved intensity based non-rigid registration results between MR scans (for example for atlas based automatic segmentation). Further validation was performed on simulated interleaving artefacts which were applied to an artefact free volume. Results obtained on prostate cancer radiotherapy treatment planning contouring were inconclusive and require further investigation.

We propose a unified variational approach for registration of gene expression data to neuroanatomical mouse atlas in two dimensions. The proposed energy (minimized in the unknown displacement u) is composed of three terms: a standard data fidelity term based on L² similarity measure, a regularizing term based on nonlinear elasticity (allowing larger smooth deformations), and a geometric penalty constraint for landmark matching. We overcome the difficulty of minimizing the nonlinear elasticity functional by introducing an auxiliary variable v that approximates ∇u, the Jacobian of the unknown displacement u. We therefore minimize now the functional with respect to the unknowns u (a vector-valued function of two dimensions) and v (a two-by-two matrix-valued function). An additional quadratic term is added, to insure good agreement between v and ∇u. In this way, the nonlinearity in the derivatives of the unknown u no longer exists in the obtained Euler-Lagrange equations, producing simpler implementations. Several satisfactory experimental results show that gene expression data are mapped to a mouse atlas with good landmark matching and smooth deformation. We also present comparisons with the biharmonic regularization. An advantage of the proposed nonlinear elasticity model is that usually no numerical correction such as regridding is necessary to keep the deformation smooth, while unifying the data fidelity term, regularization term, and landmark constraints in a single minimization approach.

In this paper, we evaluate a novel image registration method on a set of expiratory-inspiratory pairs of computed tomography (CT) lung scans. A free-form multi resolution image registration technique is used to match two scans of the same subject. To account for the differences in the lung intensities due to differences in inspiration level, we propose to adjust the intensity of lung tissue according to the local expansion or compression. An image registration method without intensity adjustment is compared to the proposed method. Both approaches are evaluated on a set of 10 pairs of expiration and inspiration CT scans of children with cystic fibrosis lung disease. The proposed method with mass preserving adjustment results in significantly better alignment of the vessel trees. Analysis of local volume change for regions with trapped air compared to normally ventilated regions revealed larger differences between these regions in the case of mass preserving image registration, indicating that mass preserving registration is better at capturing localized differences in lung deformation.

Single Plane Illumination Microscopy (SPIM) is an emerging microscopic technique that enables live imaging of large biological specimens in their entirety. By imaging the biological sample from multiple angles, SPIM has the potential to achieve isotropic resolution throughout relatively large biological specimens. For every angle, however, only a shallow section of the specimen is imaged with high resolution, whereas deeper regions appear increasingly blurred. Existing intensity-based registration techniques still struggle to robustly and accurately align images that are characterized by limited overlap and/or heavy blurring. To be able to register such images, we add sub-resolution fluorescent beads to the rigid agarose medium in which the imaged specimen is embedded. For each segmented bead, we store the relative location of its n nearest neighbors in image space as rotation-invariant geometric local descriptors. Corresponding beads between overlapping images are identified by matching these descriptors. The bead correspondences are used to simultaneously estimate the globally optimal transformation for each individual image. The final output image is created by combining all images in an angle-independent output space, using volume injection and local content-based weighting of contributing images. We demonstrate the performance of our approach on data acquired from living embryos of Drosophila and fixed adult C.elegans worms. Bead-based registration outperformed intensity-based registration in terms of computation speed by two orders of magnitude while producing bead registration errors below 1 μm (about 1 pixel). It, therefore, provides an ideal tool for processing of long term time-lapse recordings of embryonic development consisting of hundreds of time points.

In 2003, Maurer at al. [7] published a paper describing an algorithm that computes the exact distance transform in a linear time (with respect to image size) for the rectangular binary images in the k-dimensional space Rk and distance measured with respect to L_p-metric for 1 ≤ p ≤ ∞, which includes Euclidean distance L₂. In this paper we discuss this algorithm from theoretical and practical points of view. On the practical side, we concentrate on its Euclidean distance version, discuss the possible ways of implementing it as signed distance transform, and experimentally compare implemented algorithms. We also describe the parallelization of these algorithms and the computation time savings associated with such an implementation. The discussed implementations will be made available as a part of the CAVASS software system developed and maintained in our group [5]. On the theoretical side, we prove that our version of the signed distance transform algorithm, GBDT, returns, in a linear time, the exact value of the distance from the geometrically defined object boundary. We notice that, actually, the precise form of the algorithm from [7] is not well defined for L₁ and L∞ metrics and point to our complete proof (not given in [7]) that all these algorithms work correctly for the Lp-metric with 1 < p < ∞.

Nonrigid image registration is a key tool in medical imaging. Because of high degrees of freedom in nonrigid transforms, there have been many efforts to regularize the deformation based on some reasonable assumptions. Especially, motion invertibility and local tissue rigidity have been investigated as reasonable priors in image registration. There have been several papers on exploiting each constraint separately. These constraints are reasonable in respiratory motion estimation because breathing motion is invertible and there are some rigid structures such as bones. Using both constraints seems very attractive in respiratory motion registration since using invertibility prior alone usually causes bone warping in ribs. Using rigidity prior seems natural and straightforward. However, the "sliding effect" near the interface between rib cage and diaphragm makes problem harder because it is not locally invertible. In this area, invertibility and rigidity priors have opposite forces. Recently, we proposed a simple piecewise quadratic penalty that encourages the local invertibility of motions. In this work we relax this penalty function by using a Geman-type function that allows the deformation to be piecewise smooth instead of globally smooth. This allows the deformation to be discontinuous in the area of the interface between rib cage and diaphragm. With some small sacrifice of regularity, we could achieve more realistic discontinuous motion near diaphragm, better data fitting error as well as less bone warping. We applied this Geman-type function penalty only to the x- and y-direction partial derivatives of the z-direction deformation to address the sliding effect. 192 × 128 × 128 3D CT inhale and exhale images of a real patient were used to show the benefits of this new penalty method.

A novel hierarchical unbiased group-wise registration is developed to robustly transform each individual image towards a common space for atlas based analysis. This hierarchical group-wise registration approach consists of two main components, (1) data clustering to group similar images together and (2) unbiased group-wise registration to generate a mean image for each cluster. The mean images generated in the lower hierarchical level are regarded as the input images for the higher hierarchy. In the higher hierarchical level, these input images will be further clustered and then registered by using the same two components mentioned. This hierarchical bottom-up clustering and within-cluster group-wise registration is repeated until a final mean image for the whole population is formed. This final mean image represents the common space for all the subjects to be warped to in order for the atlas based analysis. Each individual image at the bottom of the constructed hierarchy is transformed towards the root node through concatenating all the intermediate displacement fields. In order to evaluate the performance of the proposed hierarchical registration in atlas based statistical analysis, comparisons were made with the conventional group-wise registration in detecting simulated brain atrophy as well as fractional anisotropy differences between neonates and 1-year-olds. In both cases, the proposed approach demonstrated improved sensitivity (higher t-scores) than the conventional unbiased registration approach.

Positron emission tomography (PET) images are acquired for many purposes, from diagnostic assessment to aiding in the development of novel therapies. Whatever the intended use, it is often necessary to distinguish between different anatomical regions within these images. Because of this, magnetic resonance images (MRIs) are generally acquired to provide an anatomical reference. This reference will only be accurate if the PET image is properly coregistered to the MRI; yet currently, a method to evaluate PET-MRI coregistration accuracy does not exist. This problem is compounded by the fact that two visually indistinguishable coregistration results can produce estimates of ligand binding that vary significantly. Therefore, the focus of this work was to develop a method that can evaluate coregistration performance based on measured ligand binding within certain regions of the coregistered PET image. The evaluation method is based on the premise that a more accurate coregistration will result in higher ligand binding in certain anatomical regions defined by the MRI. This fully automated method was able to assess coregistration results within the variance of an expert manual rater and shows promise as a possible coregistration cost function.

Automated labeling of the bronchial tree is essential for localization of airway related diseases (e.g. chronic bronchitis) and is also a useful precursor to lung-lobe labeling. We describe an automated method for registration-based labeling of a bronchial tree. The bronchial tree is segmented from a CT image using a region-growing based algorithm. The medial line of the extracted tree is then computed using a potential field based approach. The expert-labeled target (atlas) and the source bronchial trees in the form of extracted centerline point sets are brought into alignment by calculating a non-rigid thin-plate spline (TPS) mapping from the source to the target. The registration takes into account global as well as local variations in anatomy between the two images through the use of separable linear and non-linear components of the transformation; as a result it is well suited to matching structures that deviate at finer levels: namely higher order branches. The method is validated by registering together pairs of datasets for which the ground truth labels are known in advance: the labels are transferred after matching target to source and then compared with the true values. The method was tested on datasets each containing 18 branch centerpoints and 12 bifurcation locations (30 landmarks in total) annotated manually by a radiologist, where the performance was measured as the number of landmarks having the correct transfer of labels. An overall accuracy of labeling of 91.5 % was obtained in matching 23 pairs of datasets obtained from different patients.

Intra-operative imaging during neurosurgical procedures facilitates aggressive resections and potentially an increased surgical success rate compared to the traditional approach of relying purely on pre-operative data. However, acquisition of functional images like fMRI and DTI still have to be performed pre-operatively which necessitates registration to map them to the intra-operative image space. We present an elastic FEM-based registration algorithm which is tailored to register pre-operative to intra-operative images where a superficial tumor has been resected. To restrict matching of the cortical brain surface of the pre-operative image with the resected cavity in the intra-operative image, we define a weight function based on the "concavity" of the deformation field. These weights are applied to the load vector which effectively restricts the unwanted image forces around the resected area from matching the brain surface in the pre-operative image with the surface of the resected cavity. Another novelty of the proposed method is an adaptive multi-level FEM grid. After convergence of the algorithm on one level, the FEM grid is subdivided to add more degrees of freedom to the deformation around areas with a bad match. We present results from applying the algorithm on both 2D synthetic and medical image data and can show that the adaptivity of the grid both improves registration results and registration speed while the inclusion of the weighting function improves the results in the presence of resected tissue.

The ability to detect and match features across multiple views of a scene is a crucial first step in many computer vision algorithms for dynamic scene analysis. State-of-the-art methods such as SIFT and SURF perform successfully when applied to typical images taken by a digital camera or camcorder. However, these methods often fail to generate an acceptable number of features when applied to medical images, because such images usually contain large homogeneous regions with little color and intensity variation. As a result, tasks like image registration and 3D structure recovery become difficult or impossible in the medical domain. This paper presents a scale, rotation and color/illumination invariant feature detector and descriptor for medical applications. The method incorporates elements of SIFT and SURF while optimizing their performance on medical data. Based on experiments with various types of medical images, we combined, adjusted, and built on methods and parameter settings employed in both algorithms. An approximate Hessian based detector is used to locate scale invariant keypoints and a dominant orientation is assigned to each keypoint using a gradient orientation histogram, providing rotation invariance. Finally, keypoints are described with an orientation-normalized distribution of gradient responses at the assigned scale, and the feature vector is normalized for contrast invariance. Experiments show that the algorithm detects and matches far more features than SIFT and SURF on medical images, with similar error levels.

We developed an automated analysis pipeline to analyze 3D changes in ventricular morphology; it provides a highly sensitive quantitative marker of Alzheimer's disease (AD) progression for MRI studies. In the ADNI image database, we created expert delineations of the ventricles, as parametric surface meshes, in 6 brain MRI scans. These 6 images and their embedded surfaces were fluidly registered to MRI scans of 80 AD patients, 80 individuals with mild cognitive impairment (MCI), and 80 healthy controls. Surface averaging within subjects greatly reduced segmentation error. Surface-based statistical maps revealed powerful correlations between surface morphology at baseline and (1) diagnosis, (2) cognitive performance (MMSE scores), (3) depression, and (4) predicted future decline, over a 1 year interval, in 3 standard clinical scores (MMSE, global and sum-of-boxes CDR). We used a false discovery rate method (FDR) method based on cumulative probability plots to find that 40 subjects were sufficient to discriminate AD from normal groups. 60 and 119 subjects, respectively, were required to correlate ventricular enlargement with MMSE and clinical depression. Surface-based FDR, along with multi-atlas fluid registration to reduce segmentation error, will allow researchers to (1) estimate sample sizes with adequate power to detect groups differences, and (2) compare the power of mapping methods head-to-head, optimizing cost-effectiveness for future clinical trials.

We apply a recently developed automated brain segmentation method, FS+LDDMM, to brain MRI scans from Parkinson's Disease (PD) subjects, and normal age-matched controls and compare the results to manual segmentation done by trained neuroscientists. The data set consisted of 14 PD subjects and 12 age-matched control subjects without neurologic disease and comparison was done on six subcortical brain structures (left and right caudate, putamen and thalamus). Comparison between automatic and manual segmentation was based on Dice Similarity Coefficient (Overlap Percentage), L1 Error, Symmetrized Hausdorff Distance and Symmetrized Mean Surface Distance. Results suggest that FS+LDDMM is well-suited for subcortical structure segmentation and further shape analysis in Parkinson's Disease. The asymmetry of the Dice Similarity Coefficient over shape change is also discussed based on the observation and measurement of FS+LDDMM segmentation results.

Image registration is an important and active area of medical image processing. Given two images, the idea is to compute a reasonable displacement field which deforms one image such that it becomes similar to the other image. The design of an automatic registration scheme is a tricky task and often the computed displacement field has to be discarded, when the outcome is not satisfactory. On the other hand, however, any displacement field does contain useful information on the underlying images. It is the idea of this note, to utilize this information and to benefit from an even unsuccessful attempt for the subsequent treatment of the images. Here, we make use of typical vector analysis operators like the divergence and curl operator to identify meaningful portions of the displacement field to be used in a follow-up run. The idea is illustrated with the help of academic as well as a real life medical example. It is demonstrated on how the novel methodology may be used to substantially improve a registration result and to solve a difficult segmentation problem.

Image registration and segmentation are two important tasks in medical image analysis. However, the validation of algorithms for non-linear registration in particular often poses significant challenges:1, 2 Anatomical labeling based on scans for the validation of segmentation algorithms is often not available, and is tedious to obtain. One possibility to obtain suitable ground truth is to use anatomically labelled atlas images. Such atlas images are, however, generally limited to single subjects, and the displacement field of the registration between the template and an arbitrary data set is unknown. Therefore, the precise registration error cannot be determined, and approximations of a performance measure like the consistency error must be adapted. Thus, validation requires that some form of ground truth is available. In this work, an approach to generate a synthetic ground truth database for the validation of image registration and segmentation is proposed. Its application is illustrated using the example of the validation of a registration procedure, using 50 magnetic resonance images from different patients and two atlases. Three different non-linear image registration methods were tested to obtain a synthetic validation database consisting of 50 anatomically labelled brain scans.

This article describes two new algorithms that, when integrated into an existing semi-automatic virtual bone fragment reconstruction system, allow for more accurate anatomic restoration. Furthermore, they spare the user the painstaking task of positioning each fragment in 3D, which can be extremely time consuming and difficult. The virtual interactive environment gives the user capabilities to influence the reconstruction process and that allows idiosyncratic geometric surface reconstruction scenarios. Coarse correspondences specified by the user are refined by a new alignment functional that allows geometric surface variations such as ridges and valleys to more heavily influence the final alignment solution. Integration of these algorithms into the system provides improved reconstruction accuracy, which is critical for increasing the likelihood of satisfactory clinical outcome after the injury.

Endorectal MRI provides detailed images of the prostate anatomy and is useful for radiation treatment planning. The endorectal probe (which is often removed during radiotherapy) introduces a large prostate deformation, thereby posing a challenge for purposes of treatment planning. The probe-in MRI needs to be deformably registered to the planning MRI prior to radiation treatment. The goal of this paper is to evaluate a deformable registration workflow and quantify its accuracy and suitability for radiation treatment planning. We use three metrics to evaluate the accuracy of the prostate/tumor segmentations from the registered volume to the gold standard prostate/tumor segmentations: (a) Dice Similarity Coefficient (b) Hausdorff Distance (c) Mean surface distance. These metrics quantify the acceptability of the registration within the prescribed treatment margin. We evaluate and adapt existing methods, both manual and automated, to accurately track, visualize and quantify the deformations in the prostate geometry between the endorectal MRI and the treatment planning image. An important aspect of the work described in this paper is the integration of interactive guidance on the registration process. The approach described in this paper provides users with the option of performing interactive manual alignment followed by deformable registration.

The goal of radiotherapy is to deliver maximal dose to the tumor and minimal dose to the surrounding tissue. This requires accurate target definition. In sites were the tumor is difficult to see on the CT images, such as for rectal cancer, PET-CT imaging can be used to better define the target. If the information from multiple PETCT images with different tracers needs to be combined, a nonrigid registration is indispensable to compensate for rectal tissue deformations. Such registration is complicated by the presence of different volumes of bowel gas in the images to be registered. In this paper, we evaluate the performance of different nonrigid registration approaches by looking at the overlap of manually delineated rectum contours after registration. Using a B-spline transformation model, the results for two similarity measures, sum of squared differences and mutual information, either calculated over the entire image or on a region of interest are compared. Finally, we also assess the effect of the registration direction. We show that the combination of MI with a region of interest is best able to cope with residual rectal contrast and differences in bowel filling. We also show that for optimal performance the registration direction should be chosen depending on the difference in bowel filling in the images to be registered.

Verification of registration accuracy is paramount for assessing the validity and clinical feasibility of a registration method. When a ground truth registration is not available or when local misalignments need to be examined, a qualitative assessment of registration results must be performed. The verification of registration by analyzing correspondences of gradients derived from the rigidly registered CT and MR images was performed. Strongest local CT gradients were extracted and transformed into the MR gradient image. A local gradient correspondence search in the MR image was performed using discrete systematic displacements in the direction of the strongest local CT gradients. As the measure of gradient correspondence between the CT and MR gradients the absolute values and the directions of the gradients were considered. The directional information was integrated by means of a weighting function which was calculated as a product of absolute values of the strongest local CT gradient and the MR gradient weighted by the angle between these two gradients. Two correspondence visualization techniques and gradient displacement analysis were developed to highlight the misaligned gradients and provide qualitative assessment of local misregistration. The feasibility of the proposed approach was demonstrated on the CT and MR images of the RIRE database registered using the normalized mutual information similarity measure. Global and local misregistrations were detected. Furthermore, the acquisition artifacts of non-rectified MR images could be visualized and were shown to degrade the registration performance.

Fiducial-based rigid registration is the preferred method for aligning the preoperative image with the intra-operative physical anatomy in existing image-guided surgery systems. After registration, the targets locations usually cannot be measured directly, so the Target Registration Error (TRE) is often estimated with the Fiducial Registration Error (FRE), or with Fitzpatrick TRE (FTRE) estimation formula. However, large discrepancies between the FRE and the TRE have been exemplified in hypothetical setups and have been observed in the clinic. In this paper, we formally prove that in the worst case the FRE and the TRE, and the FTRE and the TRE are independent, regardless of the target location, it location, the number of fiducials, and their configuration. The worst case occurs when the unknown Fiducial Localization Error (FLE) is modeled as an affine anisotropic inhomogeneous bias. Our results generalize previous examples, contribute to the mathematical understanding of TRE estimation in fiducial-based rigid-body registration, and strengthen the need for realistic and reliable FLE models and effective TRE estimation methods.

Tensor scale (t-scale) is a local morphometric parameter describing local structure shape, orientation and scale. At any image location, t-scale is the parametric representation of the largest ellipse (an ellipsoid in 3D) centered at that location and contained in the same homogeneous region. Recently, we have improved the t-scale computation algorithm by: (1) optimizing digital representations for LoG and DoG kernels for edge detection and (2) ellipse fitting by using minimization of both algebraic and geometric distance errors. Also, t-scale has been applied to computing the deformation vector field with applications to medical image registration. Currently, the method is implemented in twodimension (2D) and the deformation vector field is directly computed from t-scale-derived normal vectors at matching locations in two images to be registered. Also, the method has been used to develop a simple algorithm for computing 2D warping from one shape onto another. Normal vector yields local structure orientation pointing to the closest edge. However, this information is less reliable along the medial axis of a shape as it may be associated with either of the two opposite edges of the local shape. This problem is overcome using a shape-linearity measure estimating relative changes in scale along the orthogonal direction. Preliminary results demonstrate the method's potential in estimating deformation between two images.

Dual modality PET-CT imaging provides aligned anatomical (CT) and functional (PET) images in a single scanning session, which can potentially be used to improve image segmentation of PET-CT data. The ability to distinguish structures for segmentation is a function of structure and modality and varies across voxels. Thus optimal contribution of a particular modality to segmentation is spatially variant. Existing segmentation algorithms, however, seldom account for this characteristic of PET-CT data and the results using these algorithms are not optimal. In this study, we propose a relative discrimination index (RDI) to characterize the relative abilities of PET and CT to correctly classify each voxel into the correct structure for segmentation. The definition of RDI is based on the information entropy of the probability distribution of the voxel's class label. If the class label derived from CT data for a particular voxel has more certainty than that derived from PET data, the corresponding RDI will have a higher value. We applied the RDI matrix to balance adaptively the contributions of PET and CT data to segmentation of brain PET-CT images on a voxel-by-voxel basis, with the aim to give the modality with higher discriminatory power a larger weight. The resultant segmentation approach is distinguished from traditional approaches by its innovative and adaptive use of the dual-modality information. We compared our approach to the non-RDI version and two commonly used PET-only based segmentation algorithms for simulation and clinical data. Our results show that the RDI matrix markedly improved PET-CT image segmentation.

This research addresses the problem of determining the location of a pulmonary nodule in a radiograph with the aid of a pre-existing computed tomographic (CT) scan. The nodule is segmented in the radiograph using a level set segmentation method that incorporates characteristics of the nodule in a digitally reconstructed radiograph (DRR) that is calculated from the CT scan. The segmentation method includes two new level set energy terms. The contrast energy seeks to increase the contrast of the segmented region relative to its surroundings. The gradient direction convergence energy is minimized when the intensity gradient direction in the region converges to a point. The segmentation method was tested on 23 pulmonary nodules from 20 cases for which both a radiographic image and CT scan were collected. The mean nodule effective diameter is 22.5 mm. The smallest nodule has an effective diameter of 12.0 mm and the largest an effective diameter of 48.1 mm. Nodule position uncertainty was simulated by randomly offsetting the true nodule center from an aim point. The segmented region is initialized to a circle centered at the aim point with a radius that is equal to the effective radius of the nodule plus a 10.0 mm margin. When the segmented region that is produced by the proposed method is used to localize the nodule, the average reduction in nodule-position uncertainty is 46%. The relevance of this method to the detection of radiotherapy targets at the time of treatment is discussed.

Pleural thickenings are caused by asbestos exposure and may evolve into malignant pleural mesothelioma. The detection of pleural thickenings is today mostly done by a visual inspection of CT data, which is time-consuming and underlies the physician's subjective judgment. We propose a new detection algorithm within our computer-assisted diagnosis (CAD) system to automatically detect pleural thickenings within CT data. First, pleura contours are identified by thresholding and contour relaxation with a probabilistic model. Subsequently, the approach to automatically detect pleural thickenings is proposed as a two-step procedure. Step one; since pleural thickenings appear as fine-scale occurrences on the rather large-scale pleura contour, a surface-based smoothing algorithm is developed. Pleural thickenings are initially detected as the difference between the original contours and the resulting "healthy" model of the pleura. Step two; as pleural thickenings can expand into the surrounding thoracic tissue, a subsequent tissue-specific segmentation for the initially detected pleural thickenings is performed in order to separate pleural thickenings from the surrounding thoracic tissue. For this purpose, a probabilistic Hounsfield model for pleural thickenings as a mixture of Gaussian distributions has been constructed. The parameters were estimated by applying the Expectation-Maximization (EM) algorithm. A model fitting technique in combination with the application of a Gibbs-Markov random field (GMRF) model then allows the tissuespecific segmentation of pleural thickenings with high precision. With these methods, a new approach is presented in order to assure a precise and reproducible detection of pleural mesothelioma in its early stage.

Abdominal CT images are commonly used for the diagnosis of kidney diseases. With the advances of CT technology, processing of CT images has become a challenging task mainly because of the large number of CT images being studied. This paper presents a texture-learning based system for the three-dimensional (3D) segmentation of renal parenchyma in abdominal CT images. The system is designed to automatically delineate renal parenchyma and is based on the texturelearning and the region-homogeneity-based approaches. The first approach is achieved with the texture analysis using the gray-level co-occurrence matrix (GLCM) features and an artificial neural network (ANN) to determine if a pixel in the CT image is likely to fall within the renal parenchyma. The second approach incorporates a two-dimensional (2D) region growing to segment renal parenchyma in single CT image slice and a 3D region growing to propagate the segmentation results to neighboring CT image slices. The criterion for the region growing is a test of region-homogeneity which is defined by examining the ANN outputs. In system evaluation, 10 abdominal CT image sets were used. Automatic segmentation results were compared with manually segmentation results using the Dice similarity coefficient. Among the 10 CT image sets, our system has achieved an average Dice similarity coefficient of 0.87 that clearly shows a high correlation between the two segmentation results. Ultimately, our system could be incorporated in applications for the delineation of renal parenchyma or as a preprocessing in a CAD system of kidney diseases.