This PDF file contains the front matter associated with SPIE Proceedings volume 7623, including the Title Page, Copyright information, Table of Contents, Introduction, and the Conference Committee listing.

Cochlear implant surgery is a procedure performed to treat profound hearing loss. Since the cochlea is not visible in surgery, the physician uses anatomical landmarks to estimate the pose of the cochlea. Research has indicated that implanting the electrode in a particular cavity of the cochlea, the scala tympani, results in better hearing restoration. The success of the scala tympani implantation is largely dependent on the point of entry and angle of electrode insertion. Errors can occur due to the imprecise nature of landmark-based, manual navigation as well as inter-patient variations between scala tympani and the anatomical landmarks. In this work, we use point distribution models of the intra-cochlear anatomy to study the inter-patient variations between the cochlea and the typical anatomic landmarks, and we implement an active shape model technique to automatically localize intra-cochlear anatomy in conventional CT images, where intra-cochlear structures are not visible. This fully automatic segmentation could aid the surgeon to choose the point of entry and angle of approach to maximize the likelihood of scala tympani insertion, resulting in more substantial hearing restoration.

Different imaging modalities or protocols of a single patient may convey different types of information regarding a disease for the same anatomical organ/tissue. On the other hand, multi-modal/multi-protocol medical images from several different patients can also provide spatial statistics of the disease occurrence, which in turn can greatly aid in disease diagnosis and aid in improved, accurate biopsy and targeted treatment. It is therefore important to not only integrate medical images from multiple patients into a common coordinate frame (in the form of a population-based atlas), but also find the correlation between these multi-modal/multi-protocol data features and the disease spatial distribution in order to identify different quantitative structural and functional disease signatures. Most previous work on construction of anatomical atlases has focused on deriving a population-based atlas for the purpose of deriving the spatial statistics. Moreover, these models are typically derived from normal or healthy subjects, either explicitly or implicitly, where it is assumed that the inter-patient pathological variation is not large. These methods are not suitable for constructing a disease atlas, where significant differences between patients on account of disease related variations can be expected. In this paper, we present a novel framework for the construction of a multi-parametric MRI-based data-driven disease atlas consisting of multi-modal and multi-protocol data from across multiple patient studies. Our disease atlas contains 3 Tesla structural (T2) and functional (dynamic contrast enhanced (DCE)) prostate in vivo MRI with corresponding whole mount histology specimens obtained via radical prostatectomy. Our atlas construction framework comprises 3 distinct modules: (a) determination of disease spatial extent on the multi-protocol MR imagery for each patient, (b) construction of a multi-protocol MR imaging spatial atlas which captures the geographical proclivity of the disease, and (c) feature extraction and the construction of the data-driven multi-protocol MRI based prostate cancer atlas. The marriage of data driven and spatial atlases could serve as a useful tool for clinicians to identifying structural and functional imaging disease signatures so as to make better, more informed diagnoses. Each spatial location in this atlas can be associated with a high dimensional multi-attribute quantitative feature vector. Additionally, since the feature vectors are extracted from across multiple patient studies, each spatial location in the data-driven atlas can be characterized by a feature distribution (in turn characterized by a mean and standard deviation). Preliminary investigation in quantitatively correlating the disease signatures from across the spatial and data driven atlases suggests that our quantitative atlas framework could emerge as a powerful tool for discovering prostate cancer imaging signatures.

Distortion of Left Ventricle (LV) external anatomy is related to some dysfunctions, such as hypertrophy. The architecture of myocardial fibers determines LV electromechanical activation patterns as well as mechanics. Thus, their joined modelling would allow the design of specific interventions (such as peacemaker implantation and LV remodelling) and therapies (such as resynchronization). On one hand, accurate modelling of external anatomy requires either a dense sampling or a continuous infinite dimensional approach, which requires non-Euclidean statistics. On the other hand, computation of fiber models requires statistics on Riemannian spaces. Most approaches compute separate statistical models for external anatomy and fibers architecture. In this work we propose a general mathematical framework based on differential geometry concepts for computing a statistical model including, both, external and fiber anatomy. Our framework provides a continuous approach to external anatomy supporting standard statistics. We also provide a straightforward formula for the computation of the Riemannian fiber statistics. We have applied our methodology to the computation of complete anatomical atlas of canine hearts from diffusion tensor studies. The orientation of fibers over the average external geometry agrees with the segmental description of orientations reported in the literature.

Computed tomography angiography (CTA), a non-invasive imaging technique, is becoming increasingly popular for cardiac examination, mainly due to its superior spatial resolution compared to MRI. This imaging modality is currently widely used for the diagnosis of coronary artery disease (CAD) but it is not commonly used for the diagnosis of ventricular and atrial function. In this paper, we present a fully automatic method for segmenting the whole heart (i.e. the outer surface of the myocardium) and cardiac chambers from CTA datasets. Cardiac chamber segmentation is particularly valuable for the extraction of ventricular and atrial functional information, such as stroke volume and ejection fraction. With our approach, we aim to improve the diagnosis of CAD by providing functional information extracted from the same CTA data, thus not requiring additional scanning. In addition, the whole heart segmentation method we propose can be used for visualization of the coronary arteries and for obtaining a region of interest for subsequent segmentation of the coronaries, ventricles and atria. Our approach is based on multi-atlas segmentation, and performed within a non-rigid registration framework. A leave-one-out quantitative validation was carried out on 8 images. The method showed a high accuracy, which is reflected in both a mean segmentation error of 1.05±1.30 mm and an average Dice coefficient of 0.93. The robustness of the method is demonstrated by successfully applying the method to 243 additional datasets, without any significant failure.

Using an atlas, an image can be segmented by mapping its coordinate space to that of the atlas in an anatomically correct way. In order to find the correct mapping between the two different coordinate spaces e.g. diffeomorphic demons registration can be applied. The demons algorithm is a popular choice for deformable image registration and offers the possibility to perform computationally efficient non-rigid (diffeomorphic) registration. However, this registration method is prone to image artifacts and image noise. Therefore it has been the main objective of the presented work to combine the efficiency of diffeomorphic demons and the stability of statistical models. In the presented approach a statistical deformation model that describes "anatomically correct" displacements vector fields for a specific registration problem is used to guide the demons registration algorithm. By projecting the current displacement vector field, which is calculated during any iteration of the registration process, into the model space a regularized version of the vector field can be computed. Using this regularized vector field for the update of the deformation field in the subsequent iteration of the registration process the demons registration algorithm can be guided by the deformation model. The proposed method was evaluated on 21 CT datasets of the right hip. Measuring the average and maximum segmentation error for all 21 datasets and all 120 test configurations it could be demonstrated that the newly proposed algorithm leads to a reduction of the segmentation error of up to 13% compared to using the conventional diffeomorphic demons algorithm.

We present a novel variational framework for deformable multi-modal image registration. Our approach is based on Renyi's statistical dependence measure of two random variables with the use of reproducing kernel Hilbert spaces associated with Gaussian kernels to simplify the computation. The popularly used method of maximizing mutual information based optimization algorithms are complex and sensitive to the quantization of the intensities, because it requires the estimation of continuous joint probability density function (pdf). The proposed model does not deal with joint pdf but instead observed independent samples. Experimental results are provided to show the effectiveness of the model.

An image, being a continuous function, is commonly discretely represented as a set of sample values, namely the intensities, associated with the spatial grids. After that, all types of the operations are then carried out there in. We denote such representation as the discrete function representation (DFR). In this paper we provide another discrete representation for images using the point sets, called the point set representation (PSR). Essentially, the image is normalized to have the unit integral and is treated as a probability density function of some random variable. Then the PSR is formed by drawing samples of such random variable. In contrast with the DFR, here the image is purely represented as points and no values are associated. Besides being an equivalent discrete representation for images, we show that certain image operations benefit from such representation in the numerical stability, performance, or both. Examples are given in the Perona-Malik type diffusion where in the PSR there is no such problem as the numerical instability. Furthermore, PSR naturally bridges the fields of image registration with the point set registration. This helps handle some otherwise difficult problems in image registration such as partial image registration, with much faster convergence speed.

In registration-based analyses of lung biomechanics and function, high quality registrations are essential to obtain meaningful results. Various criteria have been suggested to find the correspondence mappings between two lung images acquired at different levels of inflation. In this paper, we describe a new metric, the sum of squared vesselness measure difference (SSVMD), that utilizes the rich information of blood vessel locations and matches similar vesselness patterns in two images. Preserving both the lung tissue volume and the vesselness measure, a registration algorithm is developed to minimize the sum of squared tissue volume difference (SSTVD) and SSVMD together. We compare the registration accuracy using SSTVD + SSVMD with that using SSTVD alone by registering lung CT images of three normal human subjects. After adding the new SSVMD metric, the improvement of registration accuracy is observed by landmark error and fissure positioning error analyses. The average values of landmark error and fissure positioning error are reduced by about 30% and 25%, respectively. The mean landmark error is on the order of 1 mm. Statistical testing of landmark errors shows that there is a statistically significant difference between two methods with p values < 0.05 in all three subjects. Visual inspection shows there are obvious accuracy improvements in the lung regions near the thoracic cage after adding SSVMD.

In this work we evaluate a novel method for multi-modal image registration of MR images. The key feature of our approach is a new distance measure that allows for comparing modalities that are related by an arbitrary gray-value mapping. The novel measure is formulated as least square problem for minimizing the sum of squared differences of two images with respect to changing gray-values of one of the images. It turns out that the novel measure can be computed explicitly and allows for very simple and efficient implementation. We compare our new approach to rigid registration with cross-correlation, mutual information, and normalized gradient fields as distance measure.

Quadratic regularizers are used in nonparametric registration to ensure that the registration problem is well posed and to yield solutions that exhibit certain types of smoothness. Examples of popular quadratic regularizers include the diffusion, elastic, fluid, and curvature regularizers.¹ Two important features of these regularizers are whether they account for coupling of the spatial components of the deformation (elastic/fluid do; diffusion/curvature do not) and whether they are robust to initial affine misregistrations (curvature is; diffusion/ elastic/fluid are not). In this article, we show how to extend this list of quadratic regularizers to include a second-order regularizer that exhibits the best of both features: it accounts for coupling of the spatial components of the deformation and contains affine transformations in its kernel. We then show how this extended taxonomy of quadratic regularizers is related to other families of regularizers, including Cachier and Ayache's differential quadratic forms2 and Arigovindan's family of rotationally invariant regularizers.^{3, 4} Next, we describe two computationally efficient paradigms for performing nonparametric registration with the proposed regularizer, based on Fourier methods5 and on successive Gaussian convolution.^{6, 7} Finally, we illustrate the performance of the quadratic regularizers on the task of registering serial 3-D CT exams of patients with lung nodules.

A novel approach for coupling brain tumor mass effect with a continuous model of cancer progression is proposed. The purpose of the present work is to devise an efficient approximate model for the mechanical interaction of the tumor with its surroundings in order to aid registration of brain tumor images with statistical atlases as well as the generation of atlases of brain tumor disease. To model tumor progression a deterministic reaction-diffusion formalism, which describes the spatio-temporal dynamics of a coarse-grained population density of cancerous cells, is discretized on a regular grid. Tensor information obtained from a probabilistic atlas is used to model the anisotropy of the diffusion of malignant cells within white matter. To account for the expansive nature of the tumor a parametric deformation model is linked to the computed net cell density of cancerous cells. To this end, we formulate a constrained optimization problem using an inhomogeneous regularization that in turn allows for approximating physical properties of brain tissue. The described coupling model can in general be applied to estimate mass effect of non-convex, diffusive as well as multifocal tumors so that no simplification of the growth model has to be stipulated. The present work has to be considered as a proof-of-concept. Visual assessment of the computed results demonstrates the potential of the described method. We conclude that the analogy to the problem formulation in image registration potentially allows for a sensible integration of the described approach into a unified framework of image registration and tumor modeling.

In this paper we address the problem of 3D shape reconstruction from sparse X-ray projections. We present a correspondence free method to fit a statistical shape model to two X-ray projections, and illustrate its performance in 3D shape reconstruction of the femur. The method alternates between 2D segmentation and 3D shaoe reconstruction, where 2D segmentation is guided by dynamic programming along the model projection on the X-ray plane. 3D reconstruction is based on the iterative minimization of the 3D distance between a set of support points and the back-projected silhouette with respect to the pose and model parameters. We show robustness of the reconstruction on simulated X-ray projection data of the femur, varying the field of view; and in a pilot study on cadaveric femora.

In this work, we present a four-dimensional reconstruction technique for cardiac gated SPECT images using a content-adaptive deformable mesh model. Cardiac gated SPECT images are affected by a high level of noise. Noise reduction methods usually do not account for cardiac motion and therefore introduce motion blur-an artifact that can decrease diagnostic accuracy. Additionally, image reconstruction methods typically rely on uniform sampling and Cartesian griding for image representation. The proposed method utilizes a mesh representation of the images in order to utilize the benefits of content-adaptive nonuniform sampling. The mesh model allows for accurate representation of important regions while significantly compressing the data. The content-adaptive deformable mesh model is generated by combining nodes generated on the full torso using pre-reconstructed emission and attenuation images with nodes accurately sampled on the left ventricle. Ventricular nodes are further displaced according to cardiac motion using our previously introduced motion estimation technique. The resulting mesh structure is then used to perform iterative image reconstruction using a mesh-based maximum-likelihood expectation-maximization algorithm. Finally, motion-compensated post-reconstruction temporal filtering is applied in the mesh domain using the deformable mesh model. Reconstructed images as well as quantitative evaluation show that the proposed method offers improved image quality while reducing the data size.

Several studies based on biplanar radiography technologies are foreseen as great systems for 3D-reconstruction applications for medical diagnoses. This paper proposes a non-rigid registration method to estimate a 3D personalized shape of bone models from two planar x-ray images using an as-rigid-as-possible deformation approach based on a moving least-squares optimization method. Based on interactive deformation methods, the proposed technique has the ability to let a user improve readily and with simplicity a 3D reconstruction which is an important step in clinical applications. Experimental evaluations of six anatomical femur specimens demonstrate good performances of the proposed approach in terms of accuracy and robustness when compared to CT-scan.

The automatic recognition of vascular trees is a challenging task, required for roadmapping or advanced visualization. For instance, during an endovascular aneurysm repair (EVAR), the recognition of abdominal arteries in angiograms can be used to select the appropriate stent graft. This choice is based on a reduced set of arteries (aorta, renal arteries, iliac arteries) whose relative positions are quite stable. We propose in this article a recognition process based on a structural model. The centerlines of the target vessels are represented by a set of control points whose relative positions are constrained. To find their position in an angiogram, we enhance the target vessels and extract a set of possible positions for each control point. Then, a constraint propagation algorithm based on the model prunes those sets of candidates, removing inconsistent ones. We present preliminary results on 5 cases, illustrating the potential of this approach and especially its ability to handle the high variability of the target vessels.

Accurate definition of the aorta and pulmonary artery from three-dimensional (3D) multi-detector CT (MDCT) images is important for pulmonary applications. This work presents robust methods for defining the aorta and pulmonary artery in the central chest. The methods work on both contrast enhanced and no-contrast 3D MDCT image data. The automatic methods use a common approach employing model fitting and selection and adaptive refinement. During the occasional event that more precise vascular extraction is desired or the method fails, we also have an alternate semi-automatic fail-safe method. The semi-automatic method extracts the vasculature by extending the medial axes into a user-guided direction. A ground-truth study over a series of 40 human 3D MDCT images demonstrates the efficacy, accuracy, robustness, and efficiency of the methods.

In cardiac MR images the slice thickness is normally greater than the pixel size within the slices. In general, better segmentation and analysis results can be expected for isotropic high-resolution (HR) data sets. If two orthogonal data sets, e. g. short-axis (SA) and long-axis (LA) volumes are combined, an increase in resolution can be obtained. In this work we employ a super-resolution reconstruction (SRR) algorithm for computing high-resolution data sets from two orthogonal SA and LA volumes. In contrast to a simple averaging of both data in the overlapping region, we apply a maximum a posteriori approach. There, an observation model is employed for estimating an HR image that best reproduces the two low-resolution input data sets. For testing the SRR approach, we use clinical MRI data with an in-plane resolution of 1.5 mm×1.5 mm and a slice thickness of 8 mm. We show that the results obtained with our approach are superior to currently used averaging techniques. Due to the fact that the heart deforms over the cardiac cycle, we investigate further, how the replacement of a rigid registration by a deformable registration as preprocessing step improves the quality of the final HR image data. We conclude that image quality is dramatically enhanced by applying an SRR technique especially for cardiac MR images where the resolution in slice-selection direction is about five times lower than within the slices.

Tissue contrast and resolution of magnetic resonance neuroimaging data have strong impacts on the utility of the data in clinical and neuroscience tasks such as registration and segmentation. Lengthy acquisition times typically prevent routine acquisition of multiple MR tissue contrast images at high resolution, and the opportunity for detailed analysis using these data would seem to be irrevocably lost. This paper describes an example based approach using patch matching from a multiple resolution multiple contrast atlas in order to change an image's resolution as well as its MR tissue contrast from one pulse-sequence to that of another. The use of this approach to generate different tissue contrasts (T2/PD/FLAIR) from a single T1-weighted image is demonstrated on both phantom and real images.

A wide range of medical applications in clinic and research exploit images acquired by fast magnetic resonance imaging (MRI) sequences such as echo-planar imaging (EPI), e.g. functional MRI (fMRI) and diffusion tensor MRI (DT-MRI). Since the underlying assumption of homogeneous static fields fails to hold in practical applications, images acquired by those sequences suffer from distortions in both geometry and intensity. In the present paper we propose a new variational image registration approach to correct those EPI distortions. To this end we acquire two reference EPI images without diffusion sensitizing and with inverted phase encoding gradients in order to calculate a rectified image. The idea is to apply a specialized registration scheme which compensates for the characteristical direction dependent image distortions. In addition the proposed scheme automatically corrects for intensity distortions. This is done by evoking a problem dependent distance measure incorporated into a variational setting. We adjust not only the image volumes but also the phase encoding direction after correcting for patients head-movements between the acquisitions. Finally, we present first successful results of the new algorithm for the registration of DT-MRI datasets.

Arterial spin labeling (ASL) is a noninvasive MRI method that uses magnetically labeled blood to measure cerebral perfusion. Spatial resolution of ASL is relatively small and as a consequence perfusion from different tissue types is mixed in each pixel. An average ratio of gray matter (GM) to white matter (WM) blood flow is 3.2 to 1. Disregarding the partial volume effects (PVE) can thus cause serious errors of perfusion quantification. PVE also complicates spatial filtering of ASL images as apart from noise there is a spatial signal variation due to tissue partial volume. Recently, an algorithm for correcting PVE has been published by Asllani et al. It represents the measured magnetization as a sum of different tissue magnetizations weighted by their fractional volume in a pixel. With the knowledge of the partial volume obtained from a high-resolution MRI image, it is possible to separate the individual tissue contributions by linear regression on a neighborhood of each pixel. We propose an extension of this algorithm by minimizing the total-variation of the tissue specific magnetization. This makes the algorithm more flexible to local changes in perfusion. We show that this method can be used to denoise ASL images without mixing the WM and GM signal.

Medical images often consist of low-contrast objects corrupted by random noise arising in the image acquisition process. Thus, image denoising is one of the fundamental tasks required by medical imaging analysis. In this work, we investigate an adaptive denoising scheme based on the nonlocal (NL)-means algorithm for medical imaging applications. In contrast with the traditional NL-means algorithm, the proposed adaptive NL-means (ANL-means) denoising scheme has three unique features. First, it employs the singular value decomposition (SVD) method and the K-means clustering (K-means) technique for robust classification of blocks in noisy images. Second, the local window is adaptively adjusted to match the local property of a block. Finally, a rotated block matching algorithm is adopted for better similarity matching. Experimental results from both additive white Gaussian noise (AWGN) and Rician noise are given to demonstrate the superior performance of the proposed ANL denoising technique over various image denoising benchmarks in term of both PSNR and perceptual quality comparison.

Patients suffering from cerebral ischemia or subarachnoid hemorrhage, undergo a 4D (3D+time) CT Perfusion (CTP) scan to assess the cerebral perfusion and a CT Angiography (CTA) scan to assess the vasculature. The aim of our research is to extract the vascular information from the CTP scan. This requires thin-slice CTP scans that suffer from a substantial amount of noise. Therefore noise reduction is an important prerequisite for further analysis. So far, the few noise filtering methods for 4D datasets proposed in literature deal with the temporal dimension as a 4th dimension similar to the 3 spatial dimensions, mixing temporal and spatial intensity information. We propose a bilateral noise reduction method based on time-intensity profile similarity (TIPS), which reduces noise while preserving temporal intensity information. TIPS was compared to 4D bilateral filtering on 10 patient CTP scans and, even though TIPS bilateral filtering is much faster, it results in better vessel visibility and higher image quality ranking (observer study) than 4D bilateral filtering.

We present an affinity-based optimization method for nearly-automatic vessels segmentation in CTA scans. The desired segmentation is modeled as a function that minimizes a quadratic affinity-based functional. The functional incorporates intensity and geometrical vessel shape information and a smoothing constraint. Given a few user-defined seeds, the minimum of the functional is obtained by solving a single set of linear equations. The binary segmentation is then obtained by applying a user-selected threshold. The advantages of our method are that it requires fewer initialization seeds, is robust, and yields better results than existing graph-based interactive segmentation methods. Experimental results on 20 vessel segments including the carotid arteries bifurcation and noisy parts of the carotid yield a mean symmetric surface error of 0.54mm (std=0.28).

We introduce a new 3D curved tubular intensity model in conjunction with a model fitting scheme for accurate segmentation and quantification of thin vessels in 3D tomographic images. The curved tubular model is formulated based on principles of the image formation process, and we have derived an analytic solution for the model function. In contrast to previous straight models, the new model allows to accurately represent curved tubular structures, to directly estimate the local curvature by model fitting, as well as to more accurately estimate the shape and other parameters of tubular structures. We have successfully applied our approach to 3D synthetic images as well as 3D MRA and 3D CTA vascular images of the human. It turned out that we achieved more accurate segmentation results in comparison to using a straight model.

The analysis of echocardiograms, whether visual or automated, is often hampered by ultrasound artifacts which obscure the moving myocardial wall. In this study, a probabilistic framework for tracking the endocardial surface in 3D ultrasound images is proposed, which distinguishes between visible and artifact-obscured myocardium. Motion estimation of visible myocardium relies more using a local, data-driven tracker, whereas tracking of obscured myocardium is assisted by a global, statistical model of cardiac motion. To make this distinction, the expectation-maximization algorithm is applied in a stationary and dynamic frame-of-reference. Evaluation on 35 three-dimensional echocardiographic sequences shows that this artifact-aware tracker gives better results than when no distinction is made. In conclusion, the proposed tracker is able to reduce the influence of artifacts, potentially improving quantitative analysis of clinical quality echocardiograms.

The performance of the k-nearest neighborhoods (k-NN) classifier is highly dependent on the distance metric used to identify the k nearest neighbors of the query points. The standard Euclidean distance is commonly used in practice. This paper investigates the performance of k-NN classifier with respect to different adaptive metrics in the context of medical imaging. We propose using adaptive metrics such that the structure of the data is better described, introducing some unsupervised learning knowledge in k-NN. We investigated four different metrics are estimated: a theoretical metric based on the assumption that images are drawn from Brownian Image Model (BIM), the normalized metric based on variance of the data, the empirical metric is based on the empirical covariance matrix of the unlabeled data, and an optimized metric obtained by minimizing the classification error. The spectral structure of the empirical covariance also leads to Principal Component Analysis (PCA) performed on it which results the subspace metrics. The metrics are evaluated on two data sets: lateral X-rays of the lumbar aortic/spine region, where we use k-NN for performing abdominal aorta calcification detection; and mammograms, where we use k-NN for breast cancer risk assessment. The results show that appropriate choice of metric can improve classification.

In oncological therapy monitoring, the estimation of tumor growth from consecutive CT scans is an important aspect in deciding whether the given treatment is adequate for the patient. This can be done by measuring and comparing the volume of a lesion in the scans based on a segmentation. However, simply counting the voxels within the segmentation mask can lead to significant differences in the volume, if the lesion has been segmented slightly differently by various readers or in different scans, due to the limited spatial resolution of CT and due to partial volume effects. We present a novel algorithm for measuring the volume of liver metastases and lymph nodes which considers partial volume effects at the surface of a lesion. Our algorithm is based on a spatial subdivision of the segmentation. We have evaluated the algorithm on a phantom and a multi-reader study. Our evaluations have shown that our algorithm allows determining the volume more accurately even for larger slice thicknesses. Moreover, it reduces inter-observer variability of volume measurements significantly. The calculation of the volume takes 2 seconds for 50³ voxels on a single 2.66GHz Intel Core2 CPU.

Diabetic Retinopathy (DR) is one of the leading causes of blindness in the industrialized world. Early detection is the key in providing effective treatment. However, the current number of trained eye care specialists is inadequate to screen the increasing number of diabetic patients. In recent years, automated and semi-automated systems to detect DR with color fundus images have been developed with encouraging, but not fully satisfactory results. In this study we present the initial results of a new technique for the detection and localization of microaneurysms, an early sign of DR. The algorithm is based on three steps: candidates selection, the actual microaneurysms detection and a final probability evaluation. We introduce the new Radon Cliff operator which is our main contribution to the field. Making use of the Radon transform, the operator is able to detect single noisy Gaussian-like circular structures regardless of their size or strength. The advantages over existing microaneurysms detectors are manifold: the size of the lesions can be unknown, it automatically distinguishes lesions from the vasculature and it provides a fair approach to microaneurysm localization even without post-processing the candidates with machine learning techniques, facilitating the training phase. The algorithm is evaluated on a publicly available dataset from the Retinopathy Online Challenge.

In this study, clinically produced multiphase CT volumetric data sets (pre-contrast, arterial and venous enhanced phase) are drawn upon to transcend the intrinsic limitations of single phase data sets for the robust and accurate segmentation of the liver in typically challenging cases. As an initial step, all other phase volumes are registered to either the arterial or venous phase volume by a symmetric nonlinear registration method using mutual information as similarity metric. Once registered, the multiphase CT volumes are pre-filtered to prepare for subsequent steps. Under the assumption that the intensity vectors of different organs follow the Gaussian Mixture model (GMM), expectation maximization (EM) is then used to classify the multiphase voxels into different clusters. The clusters for liver parenchyma, vessels and tumors are combined together and provide the initial liver mask that is used to generate initial zeros level set. Conversely, the voxels classified as non-liver will guide the speed image of the level sets in order to reduce leakage. Geodesic active contour level set using the gradient vector flow (GVF) derived from one of the enhanced phase volumes is then performed to further evolve the liver segmentation mask. Using EM clusters as the reference, the resulting liver mask is finally morphologically post-processed to add missing clusters and reduce leakage. The proposed method has been tested on the clinical data sets of ten patients with relatively complex and/or extensive liver cancer or metastases. A 95.8% dice similarity index when compared to expert manual segmentation demonstrates the high performance and the robustness of our proposed method - even for challenging cancer data sets - and confirms the potential of a more thorough computational exploitation of currently available clinical data sets.

A novel method is presented for definition of search lines in a variety of surface segmentation approaches. The method is inspired by properties of electric field direction lines and is applicable to general-purpose n-D shapebased image segmentation tasks. Its utility is demonstrated in graph construction and optimal segmentation of multiple mutually interacting objects. The properties of the electric field-based graph construction guarantee that inter-object graph connecting lines are non-intersecting and inherently covering the entire object-interaction space. When applied to inter-object cross-surface mapping, our approach generates one-to-one and all-to-all vertex correspondent pairs between the regions of mutual interaction. We demonstrate the benefits of the electric field approach in several examples ranging from relatively simple single-surface segmentation to complex multiobject multi-surface segmentation of femur-tibia cartilage. The performance of our approach is demonstrated in 60 MR images from the Osteoarthritis Initiative (OAI), in which our approach achieved a very good performance as judged by surface positioning errors (average of 0.29 and 0.59 mm for signed and unsigned cartilage positioning errors, respectively).

Wireless Capsule Endoscopy (WCE) integrates wireless transmission with image and video technology. It has been used to examine the small intestine non invasively. Medical specialists look for signicative events in the WCE video by direct visual inspection manually labelling, in tiring and up to one hour long sessions, clinical relevant frames. This limits the WCE usage. To automatically discriminate digestive organs such as esophagus, stomach, small intestine and colon is of great advantage. In this paper we propose to use textons for the automatic discrimination of abrupt changes within a video. In particular, we consider, as features, for each frame hue, saturation, value, high-frequency energy content and the responses to a bank of Gabor filters. The experiments have been conducted on ten video segments extracted from WCE videos, in which the signicative events have been previously labelled by experts. Results have shown that the proposed method may eliminate up to 70% of the frames from further investigations. The direct analysis of the doctors may hence be concentrated only on eventful frames. A graphical tool showing sudden changes in the textons frequencies for each frame is also proposed as a visual aid to find clinically relevant segments of the video.

While geodesic active contours (GAC) have become very popular tools for image segmentation, they are sensitive to model initialization. In order to get an accurate segmentation, the model typically needs to be initialized very close to the true object boundary. Apart from accuracy, automated initialization of the objects of interest is an important pre-requisite to being able to run the active contour model on very large images (such as those found in digitized histopathology). A second limitation of GAC model is that the edge detector function is based on gray scale gradients; color images typically being converted to gray scale prior to computing the gradient. For color images, however, the gray scale gradient results in broken edges and weak boundaries, since the other channels are not exploited for the gradient determination. In this paper we present a new geodesic active contour model that is driven by an accurate and rapid object initialization scheme-weighted mean shift normalized cuts (WNCut). WNCut draws its strength from the integration of two powerful segmentation strategies-mean shift clustering and normalized cuts. WNCut involves first defining a color swatch (typically a few pixels) from the object of interest. A multi-scale mean shift coupled normalized cuts algorithm then rapidly yields an initial accurate detection of all objects in the scene corresponding to the colors in the swatch. This detection result provides the initial boundary for GAC model. The edge-detector function of the GAC model employs a local structure tensor based color gradient, obtained by calculating the local min/max variations contributed from each color channel (e.g. R,G,B or H,S,V). Our color gradient based edge-detector function results in more prominent boundaries compared to classical gray scale gradient based function. We evaluate segmentation results of our new WNCut initialized color gradient based GAC (WNCut-CGAC) model against a popular region-based model (Chan & Vese) on a total of 60 digitized histopathology images. Across a total of 60 images, the WNCut-CGAC model yielded an average overlap, sensitivity, specificity, and positive predictive value of 73%, 83%, 97%, 84%, compared to the Chan & Vese model which had corresponding values of 64%, 75%, 95%, 72%. The rapid and accurate object initialization scheme (WNCut) and the color gradient make the WNCut-CGAC scheme, an ideal segmentation tool for very large, color imagery.

An atlas provides a reference anatomic structure and an associated coordinate system. An atlas may be used in a variety of applications, including segmentation and registration, and can be used to characterize anatomy across a population. We present a method for generating an atlas of the human retina from 500 color fundus image pairs. Using color fundus image pairs, we register image pairs to obtain a larger anatomic field of view. Key retinal anatomic features are selected for atlas landmarks: disk center, fovea, and main vessel arch. An atlas coordinate system is defined based on the statistics of the landmarks. Images from the population are warped into the atlas space to produce a statistical retinal atlas which can be used for automatic diagnosis, concise indexing, semantic blending, etc.

The topogram is a 2D projection image of human body formed using a Computed Tomography (CT) scanner. It could be used to delimitate the desired scan range for further precise 3D CT scan. In this paper, we present a robust and efficient system for automatically determining scan ranges and their associated anatomical landmarks for topogram images. The system could handle the cases when only about 50% of the desired regions are visible. The robustness of our system can be attributed to three key ingredients: 1. The detection is based on a hierarchical network; 2. Network optimization is based on sequentially optimizing a set of subnetworks; 3. The detection probability is further refined based on the detection context. Extensive experiments (including external testing) on over 1000 topogram images show that our approach works robustly and efficiently even on very challenging data.

Detecting pigmented network is a crucial step for melanoma diagnosis. In this paper, we present a novel graphbased pigment network detection method that can find and visualize round structures belonging to the pigment network. After finding sharp changes of the luminance image by an edge detection function, the resulting binary image is converted to a graph, and then all cyclic sub-graphs are detected. Theses cycles represent meshes that belong to the pigment network. Then, we create a new graph of the cyclic structures based on their distance. According to the density ratio of the new graph of the pigment network, the image is classified as "Absent" or "Present". Being Present means that a pigment network is detected in the skin lesion. Using this approach, we achieved an accuracy of 92.6% on five hundred unseen images.

We proposed an approach based on reconstructive sparse representations to segment tissues in optical images of the uterine cervix. Because of large variations in image appearance caused by the changing of the illumination and specular reflection, the color and texture features in optical images often overlap with each other and are not linearly separable. By leveraging sparse representations the data can be transformed to higher dimensions with sparse constraints and become more separated. K-SVD algorithm is employed to find sparse representations and corresponding dictionaries. The data can be reconstructed from its sparse representations and positive and/or negative dictionaries. Classification can be achieved based on comparing the reconstructive errors. In the experiments we applied our method to automatically segment the biomarker AcetoWhite (AW) regions in an archive of 60,000 images of the uterine cervix. Compared with other general methods, our approach showed lower space and time complexity and higher sensitivity.

Disorders of the heart valves constitute a considerable health problem and often require surgical intervention. Recently various approaches were published seeking to overcome the shortcomings of current clinical practice,that still relies on manually performed measurements for performance assessment. Clinical decisions are still based on generic information from clinical guidelines and publications and personal experience of clinicians. We present a framework for retrieval and decision support using learning based discriminative distance functions and visualization of patient similarity with relative neighborhood graphsbased on shape and derived features. We considered two learning based techniques, namely learning from equivalence constraints and the intrinsic Random Forest distance. The generic approach enables for learning arbitrary user-defined concepts of similarity depending on the application. This is demonstrated with the proposed applications, including automated diagnosis and interventional suitability classification, where classification rates of up to 88.9% and 85.9% could be observed on a set of valve models from 288 and 102 patients respectively.

In this paper, we present a shape based segmentation methodology for magnetic resonance prostate images. We first propose a new way to represent shapes via the hyperbolic tangent of the signed distance function. This effectively corrects the drawbacks of the signed distance function and yields very reasonable results for the shape registration and learning. Secondly, under a Bayesian statistical framework, instead of computing the posterior using a uniform prior, a directional distance map is introduced in order to incorporate a priori knowledge of image content as well as the estimated center of target object. Essentially, the image is modeled as a Finsler manifold and the metric is computed out of the directional derivative of the image. Then the directional distance map is computed to suppress the posterior remote from the object center. Thirdly, in the posterior image, a localized region based cost functional is designed to drive the shape based segmentation. Such cost functional utilizes the local regional information and is robust to both image noise and remote/irrelevant disturbances. With these three major components, the entire shape based segmentation procedure is provided as a complete open source pipeline and is applied to magnetic resonance image (MRI) prostate data.

We describe a simple and robust algorithm for estimating 3D shape given a number of silhouette points obtained from two or more viewpoints and a parametric model of the shape. Our algorithm minimizes (in the least squares sense) the distances from the lines obtained by unprojecting the silhouette points to 3D to their closest silhouette points on the 3D shape. The solution is found using an iterative approach. In each iteration, we locally approximate the least squares problem with a degree-4 polynomial function. The approximate problem is solved using a nonlinear conjugate gradient solver that takes advantage of its structure to perform exact and global line searches. We tested our algorithm by applying it to reconstruct patient-specific femur shapes from simulated biplanar X-ray images.

Shape description is an important step in image analysis. The skeleton is used as a simple, compact representation of a shape. A skeleton represents the line centered in the shape and must be homotopic and one point wide. Current skeletonization algorithms compute the skeleton over several image scans, using either thinning algorithms or distance transforms. The principle of thinning is to delete points as one goes along, preserving the topology of the shape. On the other hand, the maxima of the local distance transform identifies the skeleton and is an equivalent way to calculate the medial axis. However, with this method, the skeleton obtained is disconnected so it is required to connect all the points of the medial axis to produce the skeleton. In this study we introduce a translated distance transform and adapt an existing distance driven homotopic algorithm to perform skeletonization with a single scan and thus allow the processing of unbounded images. This method is applied, in our study, on micro scanner images of trabecular bones. We wish to characterize the bone micro architecture in order to quantify bone integrity.

In this article, we propose a unified statistical framework for image segmentation with shape prior information. The approach combines an explicitely parameterized point-based probabilistic statistical shape model (SSM) with a segmentation contour which is implicitly represented by the zero level set of a higher dimensional surface. These two aspects are unified in a Maximum a Posteriori (MAP) estimation where the level set is evolved to converge towards the boundary of the organ to be segmented based on the image information while taking into account the prior given by the SSM information. The optimization of the energy functional obtained by the MAP formulation leads to an alternate update of the level set and an update of the fitting of the SSM. We then adapt the probabilistic SSM for multi-shape modeling and extend the approach to multiple-structure segmentation by introducing a level set function for each structure. During segmentation, the evolution of the different level set functions is coupled by the multi-shape SSM. First experimental evaluations indicate that our method is well suited for the segmentation of topologically complex, non spheric and multiple-structure shapes. We demonstrate the effectiveness of the method by experiments on kidney segmentation as well as on hip joint segmentation in CT images.

We present a 3D extension and validation of an intra-operative registration framework that accommodates tissue resection. The framework is based on the bijective Demons method, but instead of regularizing with the traditional Gaussian smoother, we apply an anisotropic diffusion filter with the resection modeled as a diffusion sink. The diffusion sink prevents unwanted Demon forces that originates from the resected area from diffusing into the surrounding area. Another attractive property of the diffusion sink is the resulting continuous deformation field across the diffusion sink boundary, which allows us to move the boundary of the diffusion sink without changing values in the deformation field. The area of resection is estimated by a level-set method evolving in the space of image intensity disagreements in the intra-operative image domain. A product of using the bijective Demons method is that we can also provide an accurate estimate of the resected tissue in the preoperative image space. Validation of the proposed method was performed on a set of 25 synthetic images. Our experiments show a significant improvement in accommodating resection using the proposed method compared to two other Demons based methods.

The iterative closest point (ICP) algorithm is probably the most popular algorithm for fine registration of surfaces. Among its key properties are: a simple minimization scheme, proofs of convergence as well as the easiness to modify and improve it in many ways (e.g. use of fuzzy point correspondences, incorporation of a priori knowledge, extensions to non-linear deformations, speed-up strategies, etc.) while keeping the desirable properties of the original method. However, most ICP-like registration methods suffer from the fact that they only consider the distance between the surfaces to register in the criterion to minimize, and thus are highly dependent on how the surfaces are aligned in the first place. This explains why these methods are likely to be trapped in local minima and to lead to erroneous solutions. A solution to partly alleviate this problem would consist in adding higher-order information in the criterion to minimize (e.g. normals, curvatures, etc.), but previous works along these research tracks have led to computationally intractable minimization schemes. In this paper, we propose a new way to include the point unit normals in addition to the point coordinates to derive an ICP-like scheme for non-linear registration of surfaces, and we show how to keep the properties of the original ICP algorithm. Our algorithm rests on a simple formula showing how the unit normal changes when a surface undergoes a small deformation. The use of this formula in an ICP-like algorithm is made possible by adequate implementation choices, most notably the use of a local, differentiable, parametrization of the surfaces and a locally affine deformation model using this local parametrization. Then we experimentally show the strong added value of using the unit normals in a series of controlled experiments.

A major focus of computational anatomy is to extract the most relevant information to identify and characterize anatomical variability within a group of subjects as well as between different groups. The construction of atlases is central to this effort. An atlas is a deterministic or probabilistic model with intensity variance, structural, functional or biochemical information over a population. To date most algorithms to construct atlases have been based on a single subject assuming that the population is best described by a single atlas. However, we believe that in a population with a wide range of subjects multiple atlases may be more representative since they reveal the anatomical differences and similarities within the group. In this work, we propose to use the K-means clustering algorithm to partition a set of images into several subclasses, based on a joint distance which is composed of a distance quantifying the deformation between images and a dissimilarity measured from the registration residual. During clustering, the spatial transformations are averaged rather than images to form cluster centers, to ensure a crisp reference. At the end of this algorithm, the updated centers of the k clusters are our atlases. We demonstrate this algorithm on a subset of a public available database with whole brain volumes of subjects aged 18-96 years. The atlases constructed by this method capture the significant structural differences across the group.

Robust point matching (RPM) jointly estimates correspondences and non-rigid warps between unstructured point-clouds. RPM does not, however, utilize information of the topological structure or group memberships of the data it is matching. In numerous medical imaging applications, each extracted point can be assigned group membership attributes or labels based on segmentation, partitioning, or clustering operations. For example, points on the cortical surface of the brain can be grouped according to the four lobes. Estimated warps should enforce the topological structure of such point-sets, e.g. points belonging to the temporal lobe in the two point-sets should be mapped onto each other. We extend the RPM objective function to incorporate group membership labels by including a Label Entropy (LE) term. LE discourages mappings that transform points within a single group in one point-set onto points from multiple distinct groups in the other point-set. The resulting Labeled Point Matching (LPM) algorithm requires a very simple modification to the standard RPM update rules. We demonstrate the performance of LPM on coronary trees extracted from cardiac CT images. We partitioned the point sets into coronary sections without a priori anatomical context, yielding potentially disparate labelings (e.g. [1,2,3] → [a,b,c,d]). LPM simultaneously estimated label correspondences, point correspondences, and a non-linear warp. Non-matching branches were treated wholly through the standard RPM outlier process akin to non-matching points. Results show LPM produces warps that are more physically meaningful than RPM alone. In particular, LPM mitigates unrealistic branch crossings and results in more robust non-rigid warp estimates.

3D registration of brain MRI data is vital for many medical imaging applications. However, purely intensitybased approaches for inter-subject matching of brain structure are generally inaccurate in cortical regions, due to the highly complex network of sulci and gyri, which vary widely across subjects. Here we combine a surfacebased cortical registration with a 3D fluid one for the first time, enabling precise matching of cortical folds, but allowing large deformations in the enclosed brain volume, which guarantee diffeomorphisms. This greatly improves the matching of anatomy in cortical areas. The cortices are segmented and registered with the software Freesurfer. The deformation field is initially extended to the full 3D brain volume using a 3D harmonic mapping that preserves the matching between cortical surfaces. Finally, these deformation fields are used to initialize a 3D Riemannian fluid registration algorithm, that improves the alignment of subcortical brain regions. We validate this method on an MRI dataset from 92 healthy adult twins. Results are compared to those based on volumetric registration without surface constraints; the resulting mean templates resolve consistent anatomical features both subcortically and at the cortex, suggesting that the approach is well-suited for cross-subject integration of functional and anatomic data.

The Hamilton-Jacobi equation (HJE) appears widely in applied mathematics, physics, and optimal control theory. While its analytical solution is rarely available, the numerical solver is indispensable. In this work, firstly we propose a novel numerical method, based on the fast sweeping scheme, for the static HJE. Comparing with the original fast sweeping method, our algorithm speeds up the solution up to 8 times in 3D. The efficiency is due to incorporating the ideas of the fast marching into the fast sweeping. Essentially, the sweeping origin is selected so that the sweeping direction is more consistent with the information flow direction and the regions where the two directions are against are avoided. Moreover, the successive-overrelaxation nonlinear iterative method is used for faster convergence. Secondly, we provide a complete pipeline for brain tractography, in which the proposed solver is the key component for finding the optimal fiber tracts. Besides, the pipeline contains components from orientation distribution function estimation, multiple fiber extraction to the final fiber bundle volumetric segmentation, completing the process from DW-MRI image to segmented fiber bundles. The pipeline is integrated into the publicly available software 3D Slicer. The new solver has been tested and compared with the original scheme on various types of HJEs and the tractography pipeline was tested and performed consistently on all the 12 brain DW-MRI images.

Q-ball imaging (QBI) is a diffusion imaging technique that provides unique microstructural information and outperforms diffusion tensor imaging (DTI) in areas of fiber crossings. The generalized fractional anisotropy (GFA) is a widely accepted quantitative measure of fiber integrity for QBI. In this paper we demonstrate a major drawback of the GFA occuring in crossing fiber regions: the fiber integrity is heavily underestimated here. We present a new, directional anisotropy measure and compare it to the GFA using custom-designed fiber phantoms and Monte Carlo simulations. Furthermore, we present in-vivo data of the measure. With the expense of reduced CNR, the new measure allows a better quantification of fiber bundles in crossing regions and potentially yields additional information regarding non-dominant tracts within a voxel.

Diffusion tensor imaging (DTI) is widely used to characterize tissue micro-architecture and brain connectivity. Yet DTI suffers serious limitations in regions of crossing fibers because traditional tensor techniques cannot represent multiple, independent intra-voxel orientations. Compressed sensing has been proposed to resolve crossing fibers using a tensor mixture model (e.g., Crossing Fiber Angular Resolution of Intra-voxel structure, CFARI). Although similar in spirit to deconvolution approaches, CFARI uses sparsity to stabilize estimation with limited data rather than spatial consistency or limited model order. Here, we extend the CFARI approach to resolve crossing fibers through a strictly positive, parsimonious mixture model. Together with an optimized preconditioned conjugate gradient solver, estimation error and computational burden are greatly reduced over the initial presentation. Reliable estimates of intra-voxel orientations are demonstrated in simulation and in vivo using data representative of typical, low b-value (30 directions, 700 s/mm²) clinical DTI protocols. These sequences are achievable in 5 minutes at 3 T, and the whole brain CFARI analysis is tractable for routine analysis. With these improvements, CFARI provides a robust framework for identifying intra-voxel structure with conventional DTI protocols and shows great promise in helping to resolve the crossing fiber problem in current clinical imaging studies.

Recent studies reported the development of methods for rigid registration of 2D fetal brain imaging data to correct for unconstrained fetal and maternal motion, and allow the formation of a true 3D image of conventional fetal brain anatomy from conventional MRI. Diffusion tensor imaging provides additional valuable insight into the developing brain anatomy, however the correction of motion artifacts in clinical fetal diffusion imaging is still a challenging problem. This is due to the challenging problem of matching lower signal-to-noise ratio diffusion weighted EPI slice data to recover between-slice motion, compounded by the presence of possible geometric distortions in the EPI data. In addition, the problem of estimating a diffusion model (such as a tensor) on a regular grid that takes into account the inconsistent spatial and orientation sampling of the diffusion measurements needs to be solved in a robust way. Previous methods have used slice to volume registration within the diffusion dataset. In this work, we describe an alternative approach that makes use of an alignment of diffusion weighted EPI slices to a conventional structural MRI scan which provides a geometrically correct reference image. After spatial realignment of each diffusion slice, a tensor field representing the diffusion profile is estimated by weighted least squared fitting. By qualitative and quantitative evaluation of the results, we confirm the proposed algorithm successfully corrects the motion and reconstructs the diffusion tensor field.

The functional networks, extracted from fMRI images using independent component analysis, have been demonstrated informative for distinguishing brain states of cognitive functions and neurological diseases. In this paper, we propose a novel algorithm for discriminant analysis of functional networks encoded by spatial independent components. The functional networks of each individual are used as bases for a linear subspace, referred to as a functional connectivity pattern, which facilitates a comprehensive characterization of temporal signals of fMRI data. The functional connectivity patterns of different individuals are analyzed on the Grassmann manifold by adopting a principal angle based subspace distance. In conjunction with a support vector machine classifier, a forward component selection technique is proposed to select independent components for constructing the most discriminative functional connectivity pattern. The discriminant analysis method has been applied to an fMRI based schizophrenia study with 31 schizophrenia patients and 31 healthy individuals. The experimental results demonstrate that the proposed method not only achieves a promising classification performance for distinguishing schizophrenia patients from healthy controls, but also identifies discriminative functional networks that are informative for schizophrenia diagnosis.

Automatic image analysis of magnetic resonance (MR) images of the knee is simplified by bringing the knee into a reference position. While the knee is typically put into a reference position during image acquisition, this alignment will generally not be perfect. To correct for imperfections, we propose a two-step process of bone segmentation followed by elastic tissue deformation. The approach makes use of a fully-automatic segmentation of femur and tibia from T1 and T2* images. The segmentation algorithm is based on a continuous convex optimization problem, incorporating regional, and shape information. The regional terms are included from a probabilistic viewpoint, which readily allows the inclusion of shape information. Segmentation of the outer boundary of the cortical bone is encouraged by adding simple appearance-based information to the optimization problem. The resulting segmentation without the shape alignment step is globally optimal. Standard registration is problematic for knee alignment due to the distinct physical properties of the tissues constituting the knee (bone, muscle, etc.). We therefore develop an alternative alignment approach based on a simple elastic deformation model combined with strict enforcement of similarity transforms for femur and tibia based on the obtained segmentations.

Segmentation and representation of human brain cortex from Magnetic Resonance (MR) images is an important step for visualization and analysis in many neuro imaging applications. In this paper, we propose an automatic and fast algorithm to segment the brain cortex and to represent it as a geometric surface on which analysis can be carried out. The algorithm works on T1 weighted MR brain images with extracranial tissue removed. A fuzzy clustering algorithm with a parametric bias field model is applied to assign membership values of gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) to each voxel. The cortical boundaries, namely the WM-GM and GM-CSF boundary surfaces, are extracted as iso-surfaces of functions derived from these membership functions. The central surface (CS), which traces the peak values (or ridges) of the GM membership function, is then extracted using gradient vector diffusion. Our main contribution is to provide a generic, accurate, fast, yet fully-automatic approach to (i) produce a soft segmentation of the MR brain image with intensity field correction, (ii) extract both the boundary and the center of the cortex in a surface form, where the topology and geometry can be explicitly examined, and (iii) use the extracted surfaces to model the curvy, folding cortical volume, which allows an intuitive measurement of the thickness. As a demonstration, we compute cortical thickness from the surfaces and compare the results with what has been reported in the literature. The entire process from raw MR image to cortical surface reconstruction takes on average between five to ten minutes.

This paper proposes a new Expectation-Maximization curve evolution algorithm for medical image segmentation. Traditional level set algorithms perform poorly when image information is incomplete, missing or some objects are corrupted. In such cases, statistical model-based segmentation methods are widely used since they allow object shape variations subject to shape prior constraints to overcome the incomplete or noisy information. Although matching robustly in dealing with noisy and low contrast images, the shape parameters are estimated intractably through the Maximum A Posterior (MAP) framework by using incomplete image features. In this paper, we present a statistical shape-based joint curve evolution algorithm for image segmentation based on the assumption that using hidden features of the image as missing data can simplify the estimation problem and help improve the matching performance. In our method, these hidden features are designed to be the local voxel labeling data determined based on the intensity distribution of the image and priori anatomical knowledge. Using an Expectation-Maximization formulation, both the hidden features and the object shapes can be extracted. In addition, this EM-based algorithm is applied to the joint parameter and non-parameter shape model for more accurate segmentation. Comparative results on segmenting putamen and caudate shapes in MR brain images confirm both robustness and accuracy of the proposed curve evolution algorithm.

Image labeling and parcellation are critical tasks for the assessment of volumetric and morphometric features in medical imaging data. The process of image labeling is inherently error prone as images are corrupted by noise and artifact. Even expert interpretations are subject to subjectivity and the precision of the individual raters. Hence, all labels must be considered imperfect with some degree of inherent variability. One may seek multiple independent assessments to both reduce this variability as well as quantify the degree of uncertainty. Existing techniques exploit maximum a posteriori statistics to combine data from multiple raters. A current limitation with these approaches is that they require each rater to generate a complete dataset, which is often impossible given both human foibles and the typical turnover rate of raters in a research or clinical environment. Herein, we propose a robust set of extensions that allow for missing data, account for repeated label sets, and utilize training/catch trial data. With these extensions, numerous raters can label small, overlapping portions of a large dataset, and rater heterogeneity can be robustly controlled while simultaneously estimating a single, reliable label set and characterizing uncertainty. The proposed approach enables parallel processing of labeling tasks and reduces the otherwise detrimental impact of rater unavailability.

Efficient segmentation of globally optimal surfaces in volumetric images is a central problem in many medical image analysis applications. Intra-class variance has been successfully utilized, for instance, in the Chan-Vese model especially for images without prominent edges. In this paper, we study the optimization problem of detecting a region (volume) bounded by a smooth terrain-like surface, whose intra-class variance is minimized. A novel polynomial time algorithm is developed. Our algorithm is based on the shape probing technique in computational geometry and computes a sequence of O(n) maximum flows in the derived graphs, where n is the size of the input image. Our further investigation shows that those O(n) graphs form a monotone parametric flow network, which enables to solving the optimal region detection problem in the complexity of computing a single maximum flow. The method has been validated on computer-synthetic volumetric images. Its applicability to clinical data sets was demonstrated on 20 3-D airway wall CT images from 6 subjects. The achieved results were highly accurate. The mean unsigned surface positioning error of outer walls of the tubes is 0.258 ± 0.297mm, given a voxel size of 0.39 x 0.39 x 0.6mm³.

Pulmonary fissures separate human lungs into five distinct regions called lobes. Detection of fissure is essential for localization of the lobar distribution of lung diseases, surgical planning and follow-up. Treatment planning also requires calculation of the lobe volume. This volume estimation mandates accurate segmentation of the fissures. Presence of other structures (like vessels) near the fissure, along with its high variational probability in terms of position, shape etc. makes the lobe segmentation a challenging task. Also, false incomplete fissures and occurrence of diseases add to the complications of fissure detection. In this paper, we propose a semi-automated fissure segmentation algorithm using a minimal path approach on CT images. An energy function is defined such that the path integral over the fissure is the global minimum. Based on a few user defined points on a single slice of the CT image, the proposed algorithm minimizes a 2D energy function on the sagital slice computed using (a) intensity (b) distance of the vasculature, (c) curvature in 2D, (d) continuity in 3D. The fissure is the infimum energy path between a representative point on the fissure and nearest lung boundary point in this energy domain. The algorithm has been tested on 10 CT volume datasets acquired from GE scanners at multiple clinical sites. The datasets span through different pathological conditions and varying imaging artifacts.

Segmenting the lymph node regions in head and neck CT images has been a challenging topic in the area of medical image segmentation. The method proposed herein implements an atlas-based technique constrained by an active shape model (ASM) to segment the level II, III and IV lymph nodes as one structure. A leave-one-out evaluation study performed on 15 data sets shows that the results obtained with this technique are better than those obtained with a pure atlas-based segmentation method, in particular in regions of poor contrast.

Computational atlases based on nonrigid registration have found much use in the medical imaging community. To avoid bias to any single element of the training set, there are two main approaches: using a (random) subject to serve as an initial reference and posteriorly removing bias, and a true groupwise registration with a constraint of zero average transformation for direct computation of the atlas. Major drawbacks are the possible selection of an outlier on one side, and an initialization with an invalid instance on the other. In both cases there is great potential for affecting registration performance, and producing a final average image in which the structure of interest deviates from the central anatomy of the population under study. We propose an inexpensive means of reference selection based on a groupwise correspondence measure, which avoids the selection of an outlier and is independent from the atlas construction approach that follows. Thus, it improves tractability of reference selection and robustness of automated atlas construction. We illustrate the method using a set of 20 cardiac multislice computed tomography volumes.

Large variations occur in brain anatomical structures in human populations, presenting a critical challenge to the brain mapping process. This study investigates the major impact of these variations on the performance of atlas-based segmentation. It is based on two publicly available datasets, from each of which 17 T1-weighted brain atlases were extracted. Each subject was registered to every other subject using the Morphons, a non-rigid registration algorithm. The automatic segmentations, obtained by warping the segmentation of this template, were compared with the expert segmentations using Dice index and the differences were statistically analyzed using Bonferroni multiple comparisons at significance level 0.05. The results showed that an optimum atlas for accurate segmentation of all structures cannot be found, and that the group of preferred templates, defined as being significantly superior to at least two other templates regarding the segmentation accuracy, varies significantly from structure to structure. Moreover, compared to other templates, a template giving the best accuracy in segmentation of some structures can provide highly inferior segmentation accuracy for other structures. It is concluded that there is no template optimum for automatic segmentation of all anatomical structures in the brain because of high inter-subject variation. Using a single fixed template for brain segmentation does not lead to good overall segmentation accuracy. This proves the need for multiple atlas based solutions in the context of atlas-based segmentation on human brain.

The use of atlases has been shown to be a robust method for segmentation of medical images. In this paper we explore different methods of selection of atlases for the segmentation of the quadriceps muscles in magnetic resonance (MR) images, although the results are pertinent for a wide range of applications. The experiments were performed using 103 images from the Osteoarthritis Initiative (OAI). The images were randomly split into a training set consisting of 50 images and a testing set of 53 images. Three different atlas selection methods were systematically compared. First, a set of readers was assigned the task of selecting atlases from a training population of images, which were selected to be representative subgroups of the total population. Second, the same readers were instructed to select atlases from a subset of the training data which was stratified based on population modes. Finally, every image in the training set was employed as an atlas, with no input from the readers, and the atlas which had the best initial registration, judged by an appropriate registration metric, was used in the final segmentation procedure. The segmentation results were quantified using the Zijdenbos similarity index (ZSI). The results show that over all readers the agreement of the segmentation algorithm decreased from 0.76 to 0.74 when using population modes to assist in atlas selection. The use of every image in the training set as an atlas outperformed both manual atlas selection methods, achieving a ZSI of 0.82.

In tensor-based morphometry (TBM) group-wise differences in brain structure are measured using high degreeof- freedom registration and some form of statistical test. However, it is known that TBM results are sensitive to both the registration method and statistical test used. Given the lack of an objective model of group variation is it difficult to determine a best registration method for TBM. The use of statistical tests is also problematic given the corrections required for multiple testing and the notorius difficulty selecting and intepreting signigance values. This paper presents an approach to address both of these issues by combining multiple registration methods using Dempster-Shafer Evidence theory to produce belief maps of categorical changes between groups. This approach is applied to the comparison brain morphometry in aging, a typical application of TBM, using the determinant of the Jacobian as a measure of volume change. We show that the Dempster-Shafer combination produces a unique and easy to interpret belief map of regional changes between and within groups without the complications associated with hypothesis testing.

The segmentation of brain tissue in magnetic resonance imaging (MRI) plays an important role in clinical analysis and is useful for many applications including studying brain diseases, surgical planning and computer assisted diagnoses. In general, accurate tissue segmentation is a difficult task, not only because of the complicated structure of the brain and the anatomical variability between subjects, but also because of the presence of noise and low tissue contrasts in the MRI images, especially in neonatal brain images. Fuzzy clustering techniques have been widely used in automated image segmentation. However, since the standard fuzzy c-means (FCM) clustering algorithm does not consider any spatial information, it is highly sensitive to noise. In this paper, we present an extension of the FCM algorithm to overcome this drawback, by combining information from both T1-weighted (T1-w) and T2-weighted (T2-w) MRI scans and by incorporating spatial information. This new spatially constrained FCM (SCFCM) clustering algorithm preserves the homogeneity of the regions better than existing FCM techniques, which often have difficulties when tissues have overlapping intensity profiles. The performance of the proposed algorithm is tested on simulated and real adult MR brain images with different noise levels, as well as on neonatal MR brain images with the gestational age of 39 weeks. Experimental quantitative and qualitative segmentation results show that the proposed method is effective and more robust to noise than other FCM-based methods. Also, SCFCM appears as a very promising tool for complex and noisy image segmentation of the neonatal brain.

Blind decomposition of multi-spectral fluorescent image for tumor demarcation is formulated exploiting tensorial structure of the image. First contribution of the paper is identification of the matrix of spectral responses and 3D tensor of spatial distributions of the materials present in the image from Tucker3 or PARAFAC models of 3D image tensor. Second contribution of the paper is clustering based estimation of the number of the materials present in the image as well as matrix of their spectral profiles. 3D tensor of the spatial distributions of the materials is recovered through 3-mode multiplication of the multi-spectral image tensor and inverse of the matrix of spectral profiles. Tensor representation of the multi-spectral image preserves its local spatial structure that is lost, due to vectorization process, when matrix factorization-based decomposition methods (such as non-negative matrix factorization and independent component analysis) are used. Superior performance of the tensor-based image decomposition over matrix factorization-based decompositions is demonstrated on experimental red-green-blue (RGB) image with known ground truth as well as on RGB fluorescent images of the skin tumor (basal cell carcinoma).

Recent studies show that pericardial fat is associated with vascular calcification and cardiovascular risk. The fat is imaged with Computed Tomography (CT) as part of coronary calcium scoring but it is not included in routine clinical analysis due to the lack of automatic tools for fat quantification. Previous attempts to create such an automated tool have the limitations of either assuming a preset threshold or a Gaussian distribution for fat. In order to overcome these limitations, we present a novel approach using a classification-based method to discriminate fat from other tissues. The classifier is constructed from three binary SVM classifiers trained separately for multiple tissues (fat, muscle/blood and calcium), and a specific code is assigned to each tissue type based on the number of classifiers. The decisions of these binary classifiers are combined and compared with previously determined codes using a minimum Hamming decoding distance to identify fat. We also present an improved method for detection of a compact region-of-interest around the heart to reduce the number of false positives due to neighboring organs. The proposed method UH-PFAT attained a maximum overlap of 87%, and an average overlap of 76% with expert annotations when tested on unseen data from 36 subjects. Our method can be improved by identifying additional discriminative features for fat and muscle/blood separation, or by using more advanced classification approaches such as cascaded classifiers to reduce the number of false detections.

Computer assisted diagnosis algorithms are evaluated by testing them against wide-ranging sets of images arising from real clinical conditions. Detection of the distance scale and the reference grayscale present in most ultrasound images can be used to automate the calibration of physical per-pixel distances and grayscale normalization over heterogeneously acquired ultrasound datasets. This work presents novel methods for automated detection of (i) the distance scale and the spacing between its gradations, (ii) the reference grayscale. The distance scale was detected by searching for regular peaks in the 1-D autocorrelation of image pixel columns. The grayscale bar was detected by searching for contiguous sets of columns with long sequences of monotonically changing intensity. In tests on over 1000 images the distance scale detection rate was 94.8% and the correct gradation spacing was determined 91.2% of the time. The reference grayscale detection rate was 100%. A confidence measure was also introduced to characterize the certainty of the distance scale detection. An optimal confidence threshold for flagging low-confidence results that minimizes human intervention without risk of incorrect results remaining unflagged was established through ROC curve analysis.

Automatically determining the relative position of a single CT slice within a full body scan provides several useful functionalities. For example, it is possible to validate DICOM meta-data information. Furthermore, knowing the relative position in a scan allows the efficient retrieval of similar slices from the same body region in other volume scans. Finally, the relative position is often an important information for a non-expert user having only access to a single CT slice of a scan. In this paper, we determine the relative position of single CT slices via instance-based regression without using any meta data. Each slice of a volume set is represented by several types of feature information that is computed from a sequence of image conversions and edge detection routines on rectangular subregions of the slices. Our new method is independent from the settings of the CT scanner and provides an average localization error of less than 4.5 cm using leave-one-out validation on a dataset of 34 annotated volume scans. Thus, we demonstrate that instance-based regression is a suitable tool for mapping single slices to a standardized coordinate system and that our algorithm is competitive to other volume-based approaches with respect to runtime and prediction quality, even though only a fraction of the input information is required in comparison to other approaches.

The difficult problem of identifying dominant structures in unknown data sets has been elegantly addressed recently by a non-parametric information theoretic approach, the "Jump" method. The method employs an appropriate but fixed power transformation on the distortion-rate, D(R), curve estimated by the popular K-means algorithm. Although this approach yields good results asymptotically for higher dimensional spaces, in many practical cases involving lower dimensional spaces, a transformation function with a fixed power may not find the correct model order. The work presented here develops an objective function to derive a more suitable transformation function that minimizes classification error in low dimensional data sets. In addition, a number of carefully chosen K-means seeding methods based upon proper heuristic choices have been used to enhance the detection sensitivity and to allow a more accurate estimation. The proposed method has been evaluated for a large variety of datasets and compared with the original Jump method and other well-known order estimation methods such as Minimum Description Length (MDL), Akaike Information Criteria (AIC), and Consistent Akaike Information Criteria (CAIC), demonstrating superior overall performance. Comparative results for the Wisconsin Diagnostic Breast Cancer Dataset have been included. This modified information theoretic approach to model order estimation is expected to improve and validate diagnostic classification and detection of pre-cancerous lesions. Other applications such as finding plausible number of segments in image segmentation scenarios are also possible.a

A fundamental goal of the analysis of fMRI data is to locate areas of brain activation that can differentiate various cognitive tasks. Traditionally, researchers have approached fMRI analysis through characterizing the relationship between cognitive variables and individual brain voxels. In recent years, multivariate approaches (analyze more than one voxel at once) to fMRI data analysis have gained importance. But in majority of the multivariate approaches, the voxels used for classification are selected based on prior biological knowledge or discriminating power of individual voxels. We used sequential floating forward search (SFFS) feature selection approach for selecting the voxels and applied it to distinguish the cognitive states of whether a subject is doing a reasoning or a counting task. We obtained superior classifier performance by using the sequential approach as compared to selecting the features with best individual classifier performance. We analyzed the problem of over-fitting in this extremely high dimensional feature space with limited training samples. For estimating the accuracy of the classifier, we employed various estimation methods and discussed their importance in this small sample scenario. Also we modified the feature selection algorithm by adding spatial information to incorporate the biological constraint that spatially nearby voxels tends to represent similar things.

Fully automatic imaging biomarkers may allow quantification of patho-physiological processes that a radiologist would not be able to assess reliably. This can introduce new insight but is problematic to validate due to lack of meaningful ground truth expert measurements. Rather than quantification accuracy, such novel markers must therefore be validated against clinically meaningful end-goals such as the ability to allow correct diagnosis. We present a method for automatic cartilage surface smoothness quantification in the knee joint. The quantification is based on a curvature flow method used on tibial and femoral cartilage compartments resulting from an automatic segmentation scheme. These smoothness estimates are validated for their ability to diagnose osteoarthritis and compared to smoothness estimates based on manual expert segmentations and to conventional cartilage volume quantification. We demonstrate that the fully automatic markers eliminate the time required for radiologist annotations, and in addition provide a diagnostic marker superior to the evaluated semi-manual markers.

Understanding myelination in early brain development is of clinical importance, as many neurological disorders have their origin in early cerebral organization and maturation. The goal of this work is to study a large neonate database acquired with standard MR imagery to illuminate effects of early development in MRI. 90 neonates were selected from a study of healthy brain development. Subjects were imaged via MRI postnatally. MR acquisition included high-resolution structural and diffusion tensor images. Unbiased atlases for structural and DTI data were generated and co-registered into a single coordinate frame for voxel-wise comparison of MR and DTI appearance across time. All original datasets were mapped into this frame and structural image data was additionally intensity normalized. In addition, myelinated white matter probabilistic segmentations from our neonate tissue segmentation were mapped into the same space to study how our segmentation results were affected by the changing intensity characteristics in early development Linear regression maps and p-value maps were computed and visualized. The resulting visualization of voxelswise corresponding maps of all MR and DTI properties captures early development information in MR imagery. Surprisingly, we encountered regions of seemingly decreased myelinated WM probability over time even though we expected a confident increase for all of the brain. The intensity changes in the MR images in those regions help explain this counterintuitive result. The regressional results indicate that this is an effect of intensity changes due not solely to myelination processes but also likely brain dehydration processes in early postnatal development.

Compared to region of interest based DTI analysis, voxel-based analysis gives higher degree of localization and avoids the procedure of manual delineation with the resulting intra and inter-rater variability. One of the major challenges in voxel-wise DTI analysis is to get high quality voxel-level correspondence. For that purpose, current DTI analysis tools are building on nonlinear registration algorithms that deform individual datasets into a template image that is either precomputed or computed as part of the analysis. A variety of matching criteria and deformation schemes have been proposed, but often comparative evaluation is missing. In our opinion, the use of consistent and unbiased measures to evaluate current DTI procedures is of great importance and our work presents two possible measures. Specifically, we propose the evaluation criteria generalization and specificity, originally introduced by the shape modeling community, to evaluate and compare different DTI nonlinear warping results. These measures are of indirect nature and have a population wise view. Both measures incorporate information of the variability of the registration results in the template space via a voxel-wise PCA model. Thus far, we have used these measures to evaluate our own DTI analysis procedure employing fluid-based registration on scalar DTI maps. Generalization and specificity from tensor images in the template space were computed for 8 scalar property maps. We found that for our procedure an intensity-normalized FA feature outperformed the other scalar measurements. Also, using the tensor images rather than the FA maps as a comparison frame seemed to produce more robust results.

In the last few years, clinicians have started using fiber tracking algorithms for pre- and intraoperative neurosurgical planning. In the absence of a ground truth, it is often difficult to asses the validity and precision of these algorithms. To this end, we develop a realistic DTI software model in which multiple fiber bundles and their geometrical configuration may be specified, also allowing for scenarios in which fiber bundles cross or kiss and which are common bottlenecks of fiber tracking algorithms. Partial voluming, that is the contributions of multiple tissues to a voxel, is taken into account. The model gives us the possibility to compute the diffusion-weighted signal attenuation given certain tissue and scanner parameters. On the tissue side we can model the diffusion coefficients, the principal diffusion direction and the width of the fiber bundles. On the scanner side, we can model the diffusion time, the strength and direction of the applied diffusion gradient and the width of the diffusion pulse. We also include the possibility to add noise and various artifacts such as aliasing and N/2 ghosting to the model. Having generated the model of a fiber bundle, we determine the distance between the tracked fibers and the original model, thus being able to make assertions on the accuracy of the employed fiber tracking algorithm. Moreover, we can use this information to give an indication about an appropriate width of a safety margin around the tracked fiber bundle.

Fiber clustering is a very important step towards tract-based, quantitative analysis of white matter via diffusion tensor imaging (DTI). This work proposes a new computational framework for white matter fiber clustering based on symbolic sequence analysis method. We first perform brain tissue segmentation on the DTI image using a multi-channel fusion method and parcellate the whole brain into anatomically labeled regions via a hybrid volumetric and surface warping algorithm. Then, we perform standard fiber tractography on the DTI image and encode each tracked fiber by a sequence of labeled brain regions. Afterwards, the similarity between any pair of anatomically encoded fibers is defined as the similarity of symbolic sequences, which is a well-studied problem in the bioinformatics domain such as is used for gene and protein symbolic sequences comparisons. Finally, the normalized graph cut algorithm is applied to cluster the fibers into bundles based on the above defined similarities between any pair of fibers. Our experiments show promising results of the proposed fiber clustering framework.

Diffusion tensor magnetic resonance imaging (DT-MRI) is increasingly used in clinical research and applications for its ability to depict white matter tracts and for its sensitivity to microstructural and architectural features of brain tissue. However, artifacts are common in clinical DT-MRI acquisitions. Signal perturbations produced by such artifacts can be severe and neglecting to account for their contribution can result in erroneous diffusion tensor values. The Robust Estimation of Tensors by Outlier Rejection (RESTORE) has been demonstrated to be an effective method for improving tensor estimation on a voxel-by-voxel basis in the presence of artifactual data points in diffusion weighted images. Despite the very good performance of the RESTORE algorithm, there are some limitations and opportunities for improvement. Instabilities in tensor estimation using RESTORE have been observed in clinical human brain data. Those instabilities can come from the intrinsic high frequency spin inflow effects in non-DWIs or from excluding too many data points from the fitting. This paper proposes several practical constraints to the original RESTORE method. Results from Monte Carlo simulation indicate that the improved RESTORE method reduces the instabilities in tensor estimation observed from the original RESTORE method.

Schizophrenia is a serious and disabling mental disorder. Diffusion tensor imaging (DTI) studies performed on schizophrenia have demonstrated white matter degeneration either due to loss of myelination or deterioration of fiber tracts although the areas where the changes occur are variable across studies. Most of the population based studies analyze the changes in schizophrenia using scalar indices computed from the diffusion tensor such as fractional anisotropy (FA) and relative anisotropy (RA). The scalar measures may not capture the complete information from the diffusion tensor. In this paper we have applied the RADTI method on a group of 9 controls and 9 patients with schizophrenia. The RADTI method converts the tensors to log-Euclidean space where a linear regression model is applied and hypothesis testing is performed between the control and patient groups. Results show that there is a significant difference in the anisotropy between patients and controls especially in the parts of forceps minor, superior corona radiata, anterior limb of internal capsule and genu of corpus callosum. To check if the tensor analysis gives a better idea of the changes in anisotropy, we compared the results with voxelwise FA analysis as well as voxelwise geodesic anisotropy (GA) analysis.

Fiber tracking (FT) and quantification algorithms are approximations of reality due to limited spatial resolution, model assumptions, user-defined parameter settings, and physical imaging artifacts resulting from diffusion sequences. Until now, correctness, plausibility, and reliability of both FT and quantification techniques have mainly been verified using histologic knowledge and software or hardware phantoms. Probabilistic FT approaches aim at visualizing the uncertainty present in the data by incorporating models of the acquisition process and noise. The uncertainty is assessed by tracking many possible paths originating from a single seed point, thereby taking the tensor uncertainty into account. Based on the tracked paths, maps of connectivity probabilities can be produced, which may be used to delineate risk structures for presurgical planning. In this paper, we explore the advantages and disadvantages of probabilistic approaches compared to deterministic algorithms and give both qualitative and quantitative comparisons based on clinical data. We focus on two important clinical applications, namely, on the reconstruction of fiber bundles within the proximity of tumors and on the quantitative analysis of diffusion parameters along fiber bundles. Our results show that probabilistic FT is superior and suitable for a better reconstruction at the borders of anatomical structures and is significantly more sensitive than the deterministic approach for quantification purposes. Furthermore, we demonstrate that an alternative tracking approach, called variational noise tracking, is qualitatively comparable with a standard probabilistic method, but is computationally less expensive, thus, enhancing its appeal for clinical applications.

We present a novel and efficient approach for 3D motion estimation of keratin intermediate filaments in vitro. Keratin filaments are elastic cables forming a complex scaffolding within epithelial cells. To understand the mechanisms of filament formation and network organisation under physiological and pathological conditions, quantitative measurements of dynamic network alterations are essential. Therefore we acquired time-lapse series of 3D images using a confocal laser scanning microscope. Based on these image series, we show that a dense vector field can be computed such that the displacements from one frame to the next can be determined. Our method is based on a two-step registration process: First, a rigid pre-registration is applied in order to compensate for possible global cell movement. This step enables the subsequent nonrigid registration to capture only the sought local deformations of the filaments. As the transformation model of the deformable registration algorithm is based on Free Form Deformations, it is well suited for modeling filament network dynamics. The optimization is performed using efficient linear programming techniques such that the huge amount of image data of a time series can be efficiently processed. The evaluation of our results illustrates the potential of our approach.

Tags are non-invasive features induced in the heart muscle that enable the tracking of heart motion. Each tag line, in fact, corresponds to a 3D tag surface that deforms with the heart muscle during the cardiac cycle. Tracking of tag surfaces deformation is useful for the analysis of left ventricular motion. Cardiac material markers (Kerwin et al, MIA, 1997) can be obtained from the intersections of orthogonal surfaces which can be reconstructed from short- and long-axis tagged images. The proposed method uses Harmonic Phase (HARP) method for tracking tag lines corresponding to a specific harmonic phase value and then the reconstruction of grid tag surfaces is achieved by a Delaunay triangulation-based interpolation for sparse tag points. Having three different tag orientations from short- and long-axis images, the proposed method showed the deformation of 3D tag surfaces during the cardiac cycle. Previous work on tag surface reconstruction was restricted for the "dark" tag lines; however, the use of HARP as proposed enables the reconstruction of isosurfaces based on their harmonic phase values. The use of HARP, also, provides a fast and accurate way for tag lines identification and tracking, and hence, generating the surfaces.

This research aims to develop a methodological framework based on a data driven approach known as particle filters, often found in computer vision methods, to correct the effect of respiratory motion on Nuclear Medicine imaging data. Particles filters are a popular class of numerical methods for solving optimal estimation problems and we wish to use their flexibility to make an adaptive framework. In this work we use the particle filter for estimating the deformation of the internal organs of the human torso, represented by X, over a discrete time index k. The particle filter approximates the distribution of the deformation of internal organs by generating many propositions, called particles. The posterior estimate is inferred from an observation Z_k of the external torso surface. We demonstrate two preliminary approaches in tracking organ deformation. In the first approach, X_k represent a small set of organ surface points. In the second approach, X_k represent a set of affine organ registration parameters to a reference time index r. Both approaches are contrasted to a comparable technique using direct mapping to infer X_k from the observation Z_k. Simulations of both approaches using the XCAT phantom suggest that the particle filter-based approaches, on average performs, better.

Tracking the swimming motion of C. elegans worms is of high interest for a variety of research projects on behavior in biology, from aging to mating studies. We compare six different tracking methods, derived from two types of image preprocessing, namely local and global thresholding methods, and from three types of segmentation methods: low-level vision, and articulated models of either constant or varying width. All these methods have been successfully used in recent related works, with some modifications to adapt them to swimming motions. We show a quantitative comparison of these methods using computer-vision measures. To discuss their relative strengths and weaknesses, we consider three scenarios of behavior studies, depending on the constraints of a C. elegans project, and give suggestions as to which methods are more adapted to each case, and how to further improve them.

This paper describes a method of using a tracking system to track the upper part of the anterior surface during scanning for developing patient-specific models of respiration. In the experimental analysis, the natural variation in the anterior surface during breathing will be modeled to reveal the dominant pattern in the breathing cycle. The main target is to produce a patient-specific set of parameters that describes the configuration of the anterior surface for all respiration phases. These data then will be linked to internal organ motion to identify the effect of the morphology of each on motion using particle filter to account for previously unseen patterns of motion. In this initial study, a set of volunteers were imaged using the Codamotion infrared marker-based system. In the marker-based system, the temporal variation of the respiratory motion was studied. This showed that for the 12 volunteer cohort, the mean displacement of the thorax surface TS (abdomen surface AS) region is 10.7±5.6 mm (16.0±9.5mm). Finally, PCA was shown to capture the redundancy in the data set with the first principal component (PC) accounting for more than 96% of the overall variance in both AS and TS datasets. A fitting to the dominant modes of variation using a simple piecewise sinusoid has suggested a maximum error of about 1.1mm across the complete cohort dataset.

Images reconstructed from tomographic projection data are subject to motion artifacts from organs that move during the duration of the scan. The effect can be reduced by taking the motion into account in the reconstruction algorithm if an estimate of the deformation exists. This paper presents the estimation of the three-dimensional cardiac motion by registering reconstructed images from cardiac quiet phases as a first step towards motion-compensated cardiac image reconstruction. The non-rigid deformations of the heart are parametrized on a coarse grid on the image volume and are interpolated with cubic b-splines. The optimization problem of finding b-spline coefficients that best describe the observed deformations is ill-posed due to the large number of parameters and the resulting motion vector field is sensitive to the choice of initial parameters. Particularly challenging is the task to capture the twisting motion of the heart. The motion vector field from a dynamic computer phantom of the human heart is used to initialize the transformation parameters for the optimization process with realistic starting values. The results are evaluated by comparing the registered images and the obtained motion vector field to the case when the registration is performed without using prior knowledge about the expected cardiac motion. We find that the registered images are similar for both approaches, but the motion vector field obtained from motion estimation initialized with the phantom describes the cardiac contraction and twisting motion more accurately.

In this paper, we present an original method to evaluate the deformations in the third cerebral ventricle on a brain cine- MR imaging. First, a segmentation process, based on a fractional differentiation method, is directly applied on a 2D+t dataset to detect the contours of the region of interest (i.e. lamina terminalis). Then, the successive segmented contours are matched using a procedure of global alignment, followed by a morphing process, based on the Covariance Matrix Adaptation Evolution Strategy (CMAES). Finally, local measurements of deformations are derived from the previously determined matched contours. The validation step is realized by comparing our results with the measurements achieved on the same patients by an expert.

Computer assistance in Minimally Invasive Surgery is a very active field of research. Many systems designed for Computer Assisted Surgery require information about the instruments' positions and orientations. Our main focus lies on tracking a laparoscopic ultrasound probe to generate 3D ultrasound volumes. State-of-the-art tracking methods such as optical or electromagnetic tracking systems measure pose with respect to a fixed extra-body coordinate system. This causes inaccuracies of the reconstructed ultrasound volume in the case of patient motion, e.g. due to respiration. We propose attaching an endoscopic camera to the ultrasound probe and calculating the camera motion from the video sequence with respect to the organ surface. We adapt algorithms developed for solving the relative pose problem to recreate the camera path during the ultrasound sweep over the organ. By this image-based motion estimation camera motion can only be determined up to an unknown scale factor, known as the depth-speed-ambiguity. We show, how this problem can be overcome in the given scenario, exploiting the fact, that the distance of the camera to the organ surface is fixed and known. Preprocessing steps are applied to compensate for endoscopic image quality deficiencies.

The demons algorithm is a fast non-parametric non-rigid registration method. In recent years great efforts have been made to improve the approach; the state of the art version yields symmetric inverse-consistent largedeformation diffeomorphisms. However, only limited work has explored inter-modal similarity metrics, with no practical evaluation on multi-modality data. We present a diffeomorphic demons implementation using the analytical gradient of Normalised Mutual Information (NMI) in a conjugate gradient optimiser. We report the first qualitative and quantitative assessment of the demons for inter-modal registration. Experiments to spatially normalise real MR images, and to recover simulated deformation fields, demonstrate (i) similar accuracy from NMI-demons and classical demons when the latter may be used, and (ii) similar accuracy for NMI-demons on T1w-T1w and T1w-T2w registration, demonstrating its potential in multi-modal scenarios.

Evaluating non-rigid image registration algorithm performance is a difficult problem since there is rarely a "gold standard" (i.e., known) correspondence between two images. This paper reports the analysis and comparison of five non-rigid image registration algorithms using the Non-Rigid Image Registration Evaluation Project (NIREP) (www.nirep.org) framework. The NIREP framework evaluates registration performance using centralized databases of well-characterized images and standard evaluation statistics (methods) which are implemented in a software package. The performance of five non-rigid registration algorithms (Affine, AIR, Demons, SLE and SICLE) was evaluated using 22 images from two NIREP neuroanatomical evaluation databases. Six evaluation statistics (relative overlap, intensity variance, normalized ROI overlap, alignment of calcarine sulci, inverse consistency error and transitivity error) were used to evaluate and compare image registration performance. The results indicate that the Demons registration algorithm produced the best registration results with respect to the relative overlap statistic but produced nearly the worst registration results with respect to the inverse consistency statistic. The fact that one registration algorithm produced the best result for one criterion and nearly the worst for another illustrates the need to use multiple evaluation statistics to fully assess performance.

We are interested in establishing ground truth data for validating morphology measurements of human knee cartilage from MR imaging. One promising approach is to compare the high-accuracy 3D laser scans of dissected cadaver knees before and after the dissolution of their cartilage. This requires an accurate and reliable method to fuse the individual laser scans from multiple views of the cadaver knees. Unfortunately existing methods using Iterative Closest Point (ICP) algorithm from off-the-shell packages often yield unreliable fusion results. We identify two major sources of variation: (i) the noise in depth measurements of the laser scans is significantly high and (ii) the use of point-to-point correspondence in ICP is not suitable due to sampling variation in the laser scans. We resolve the first problem by performing adaptive Gaussian smoothing on each individual laser scans prior to the fusion. For the second problem, we construct a surface mesh from the point cloud of each scan and adopt a point-to-mesh ICP scheme for pairwise alignment. The complete surface mesh is constructed by fusing all the scans in the order maximizing mutual overlaps. In experiments on 6 repeated scanning trials of a cadaver knee, our approach reduced the alignment error of point-to-point ICP by 30% and reduced coefficient of variation (CV) of cartilage thickness measurements from 5% down to 1.4%, significantly improving the method's repeatability.

Clinical studies on atherosclerosis agree that multi-contrast MRI is the most promising technique for in-vivo characterization of carotid plaques. Multi-contrast image registration is essential for this application, because it corrects misalignments caused by patient motion during MRI acquisition. To date, it has not been determined which automatic method provides the best registration accuracy in carotid MRI. This study tries to answer this question by presenting an iterative coarse-to-fine algorithm that co-registers multi-contrast images of carotid arteries using three similarity metrics: Correlation Ratio (CR), Mutual Information (MI) and Gradient MI (GMI). The registration algorithm is first applied on the entire images and then only on the Region of Interest (ROI) of the carotid arteries using sub-pixel accuracy. The ROI is defined by an automatic carotid detection algorithm, which was tested on a group of 20 patients with different types of atherosclerotic plaques (sensitivity 91% and specificity 88%). Automatic registration was compared with image alignment obtained by manual operators (clinically qualified vascular specialists). Registration accuracies were measured using a novel MRI validation procedure, in which the gold standard is represented by in-plane rigid transformations applied by the MRI system to mimic neck movements. Overall, automatic methods (GMI = 181 ± 104 μm) produced lower registration errors than manual operators (365 ± 102 μm). GMI performed slightly better than CR and MI, suggesting that anatomical information improves registration accuracy in the carotid ROI.

This paper describes the development of a cylindrical affine transformation model for image registration. The usefulness of the model for initial alignment was demonstrated for the application of registering prone and supine 3D MR images of the breast. Final registration results visually improved when using the cylindrical affine transformation model instead of none or a Cartesian affine transformation model before non-rigid registration.

3D shape context is a method to define matching points between similar shapes as a pre-processing step to non-rigid registration. The main limitation of the approach is point mismatching, which includes long geodesic distance mismatch and neighbors crossing mismatch. In this paper, we propose a topological structure verification method to correct the long geodesic distance mismatch and a correspondence field smoothing method to correct the neighbors crossing mismatch. A robust 3D shape context model is proposed and further combined with thin-plate spline model for non-rigid surface registration. The method was tested on phantoms and rat hind limb skeletons from micro CT images. The results from experiments on mouse hind limb skeletons indicate that the approach is robust.

Nonlinear image registration is an initial step for a large number of medical image analysis applications. Optical flow based intensity registration is often used for dealing with intra-modality applications involving motion differences. In this work we present an energy functional which uses a novel, second-order regularization prior of the displacement field. Compared to other methods our scheme is robust to non-Gaussian noise and does not penalize locally affine deformation fields in homogeneous areas. We propose an efficient and stable numerical scheme to find the minimizer of the presented energy. We implemented our algorithm using modern consumer graphics processing units and thereby increased the execution performance dramatically. We further show experimental evaluations on clinical CT thorax data sets at different breathing states and on dynamic 4D CT cardiac data sets.

Image registration is a procedure to spatially align two images that is often used in, for example, computer-aided diagnosis or segmentation applications. To maximize the flexibility of image registration methods, they depend on many registration parameters that must be fine-tuned for each specific application. Tuning parameters is a time-consuming task, that would ideally be performed for each individual registration. However, doing this manually for each registration is too time-consuming, and therefore we would like to do this automatically. This paper proposes a methodology to estimate one of most important parameters in a registration procedure, the regularization setting, on the basis of the image similarity. We test our method on a set of images of prostate cancer patients and show that using the proposed methodology, we can improve the result of image registration when compared to using an average-best parameter.

Contrast-enhanced liver MR image sequences acquired at multiple times before and after contrast administration have been shown to be critically important for the diagnosis and monitoring of liver tumors and may be used for the quantification of liver inflammation and fibrosis. However, over multiple acquisitions, the liver moves and deforms due to patient and respiratory motion. In order to analyze contrast agent uptake one first needs to correct for liver motion. In this paper we present a method for the motion correction of dynamic contrastenhanced liver MR images. For this purpose we use a modified version of the Local Correlation Coefficient (LCC) Demons non-rigid registration method. Since the liver is nearly incompressible its displacement field has small divergence. For this reason we add a divergence term to the energy that is minimized in the LCC Demons method. We applied the method to four sequences of contrast-enhanced liver MR images. Each sequence had a pre-contrast scan and seven post-contrast scans. For each post-contrast scan we corrected for the liver motion relative to the pre-contrast scan. Quantitative evaluation showed that the proposed method improved the liver alignment relative to the non-corrected and translation-corrected scans and visual inspection showed no visible misalignment of the motion corrected contrast-enhanced scans and pre-contrast scan.

The human brain undergoes significant changes in the first few years after birth, but knowledge about this critical period of development is quite limited. Previous neuroimaging studies have been mostly focused on morphometric measures such as volume and shape, although tissue property measures related to the degree of myelination and axon density could also add valuable information to our understanding of brain maturation. Our goal is to complement brain growth analysis via morphometry with the study of longitudinal tissue property changes as reflected in patterns observed in multi-modal structural MRI and DTI. Our preliminary study includes eight healthy pediatric subjects with repeated scans at the age of two weeks, one year, and two years with T1, T2, PD, and DT MRI. Analysis is driven by the registration of multiple modalities and time points within and between subjects into a common coordinate frame, followed by image intensity normalization. Quantitative tractography with diffusion and structural image parameters serves for multi-variate tissue analysis. Different patterns of rapid changes were observed in the corpus callosum and the posterior and anterior internal capsule, structures known for distinctly different myelination growth. There are significant differences in central versus peripheral white matter. We demonstrate that the combined longitudinal analysis of structural and diffusion MRI proves superior to individual modalities and might provide a better understanding of the trajectory of early neurodevelopment.

In this paper, we propose a fast rigid-registration method of inferior limb X-ray films (two-dimensional Computed Radiography (CR) images) and three-dimensional Computed Tomography (CT) images for Total Knee Arthroplasty (TKA) surgery planning. The position of the each bone, such as femur and tibia (shin bone), in X-ray film and 3D CT images is slightly different, and we must pay attention how to use the two different images, since X-ray film image is captured in the standing position, and 3D CT is captured in decubitus (face up) position, respectively. Though the conventional registration mainly uses cross-correlation function between two images,and utilizes optimization techniques, it takes enormous calculation time and it is difficult to use it in interactive operations. In order to solve these problems, we calculate the center line (bone axis) of femur and tibia (shin bone) automatically, and we use them as initial positions for the registration. We evaluate our registration method by using three patient's image data, and we compare our proposed method and a conventional registration, which uses down-hill simplex algorithm. The down-hill simplex method is an optimization algorithm that requires only function evaluations, and doesn't need the calculation of derivatives. Our registration method is more effective than the downhill simplex method in computational time and the stable convergence. We have developed the implant simulation system on a personal computer, in order to support the surgeon in a preoperative planning of TKA. Our registration method is implemented in the simulation system, and user can manipulate 2D/3D translucent templates of implant components on X-ray film and 3D CT images.

Stable features under simulated mammographic compressions, which will become candidate landmarks for a temporal mammographic feature-based registration algorithm, are discussed in this paper. Using these simulated mammograms, we explore the extraction of features based on standard intensity projection images and local phase projection images. One approach to establishing corresponding features is by template matching using a similarity measure. Simulated mammographic projections from deformed MR volumes are employed, as the mean projected 3D displacements are computed and therefore validation of the technique is possible. Tracking is done by template matching using normalized cross correlation as the similarity measure. The performance of standard projection images and local phase projection images is compared. The preliminary results reveal that although the majority of the points within the breast are difficult to track, a small number may be successfully tracked, which is indicative of their stability and thus their suitability as candidate landmarks. Whilst matching using the standard projection images achieves an overall error of 14.46mm, this error increases to 22.7mm when computing local phase of the projection images. These results suggest that using local phase alone does not improve template matching. For the identification of stable landmarks for feature-based mammogram registration, we conclude that intensity based template matching using normalized correlation is a feasible approach for identifying stable features.

Robust and automatic non-rigid registration depends on many parameters that have not yet been systematically explored. Here we determined how tissue classification influences non-linear fluid registration of brain MRI. Twin data is ideal for studying this question, as volumetric correlations between corresponding brain regions that are under genetic control should be higher in monozygotic twins (MZ) who share 100% of their genes when compared to dizygotic twins (DZ) who share half their genes on average. When these substructure volumes are quantified using tensor-based morphometry, improved registration can be defined based on which method gives higher MZ twin correlations when compared to DZs, as registration errors tend to deplete these correlations. In a study of 92 subjects, higher effect sizes were found in cumulative distribution functions derived from statistical maps when performing tissue classification before fluid registration, versus fluidly registering the raw images. This gives empirical evidence in favor of pre-segmenting images for tensor-based morphometry.

The estimation of respiratory motion is a fundamental requisite for many applications in the field of 4D medical imaging, for example for radiotherapy of thoracic and abdominal tumors. It is usually done using non-linear registration of time frames of the sequence without further modelling of physiological motion properties. In this context, the accurate calculation of liver und lung motion is especially challenging because the organs are slipping along the surrounding tissue (i.e. the rib cage) during the respiratory cycle, which leads to discontinuities in the motion field. Without incorporating this specific physiological characteristic, common smoothing mechanisms cause an incorrect estimation along the object borders. In this paper, we present an extended diffusion-based model for incorporating physiological knowledge in image registration. By decoupling normal- and tangential-directed smoothing, we are able to estimate slipping motion at the organ borders while preventing gaps and ensuring smooth motion fields inside. We evaluate our model for the estimation of lung and liver motion on the basis of publicly accessible 4D CT and 4D MRI data. The results show a considerable increase of registration accuracy with respect to the target registration error and a more plausible motion estimation.

This paper presents a novel approach to X-ray mammography - MRI registration. The proposed method uses an intensity-based technique and an affine transformation matrix to approximate the 3D deformation of the breast resulting from the compression applied during mammogram acquisition. The registration is driven by a similarity measure that is calculated at each iteration of the algorithm between the target X-ray mammogram and a simulated X-ray image, created from the MR volume. Although the similarity measure is calculated in 2D, we compute a 3D transformation that is updated at each iteration. We have performed two types of experiments. In the first set, we used simulated X-ray target data, for which the ground truth deformation of the volume was known and thus the results could be validated. For this case, we examined the performance of 4 different similarity measures and we show that Normalized Cross Correlation and Gradient Difference perform best. The calculated mean reprojection error was for both similarity measures 4mm, for an initial misregistration of 14mm. In the second set of experiments, we present the initial results of registering real X-ray mammograms with MR volumes. The results indicate that the breast boundaries were registered well and the volume was deformed in 3D in a similar way to the deformation of the breast during X-ray mammogram acquisition. The experiments were carried out on five patients.

Stitching of volumes obtained from three dimensional (3D) ultrasound (US) scanners improves visualization of anatomy in many clinical applications. Fast but accurate volume registration remains the key challenge in this area.We propose a volume stitching method based on efficient registration of 3D US volumes obtained from a tracked US probe. Since the volumes, after adjusting for probe motion, are coarsely registered, we obtain salient correspondence points in the central slices of these volumes. This is done by first removing artifacts in the US slices using intensity invariant local phase image processing and then applying the Harris Corner detection algorithm. Fast sub-volume registration on a small neighborhood around the points then gives fast, accurate 3D registration parameters. The method has been tested on 3D US scans of phantom and real human radius and pelvis bones and a phantom human fetus. The method has also been compared to volumetric registration, as well as feature based registration using 3D-SIFT. Quantitative results show average post-registration error of 0.33mm which is comparable to volumetric registration accuracy (0.31mm) and much better than 3D-SIFT based registration which failed to register the volumes. The proposed method was also much faster than volumetric registration (~4.5 seconds versus 83 seconds).