Information Processing in Medical Imaging

We did human observer experiments using a hybrid image technique to determine the variation of tumor contrast thresholds for detection as a function of tumor sizes. This was done with both mammographic backgrounds and filtered noise with the same power spectra. We obtained the very surprising result that contrast had to be increased as lesion size increased to maintain contrast detectability. All previous investigations with white noise, radiographic and CT imaging system noise have shown the opposite effect. We compared human results to predictions of a number of observer models and found fairly good qualitative agreement. However we found that human performance was better than what would be expected if mammographic structure was assumed to be pure noise. This disagreement can be accounted for by using a simple scaling correction factor.

Imaging is often used for the purpose of estimating the value of some parameter of interest. For example, a cardiologist may measure the ejection fraction (EF) of the heart in order to know how much blood is being pumped out of the heart on each stroke. In clinical practice, however, it is difficult to evaluate an estimation method because the gold standard is not known, e.g., a cardiologist does not know the true EF of a patient. Thus, researchers have often evaluated an estimation method by plotting its results against the results of another (more accepted) estimation method, which amounts to using one set of estimates as the pseudogold standard. In this paper, we present a maximum likelihood approach for evaluating and comparing different estimation methods without the use of a gold standard with specific emphasis on the problem of evaluating EF estimation methods. Results of numerous simulation studies will be presented and indicate that the method can precisely and accurately estimate the parameters of a regression line without a gold standard, i.e., without the x-axis.

This paper presents detailed derivations of an unbiased estimate for an observer template (a set of linear pixel weights an observer uses to perform a visual task) in two-alternative forced-choice experiments. Two derivations of the covariance matrix associated with the error present in this estimation method are also derived and compared in human-observer data.

Continuum mechanical models have been used to regularize ill-posed problems in many applications in medical imaging analysis such as image registration and left ventricular motion estimation. In this work, we present a significant extension to the common elastic model which we call the active elastic model. The active elastic model is designed to reduce bias in deformation estimation and to allow the imposition of proper priors on deformation estimation problems that contain information regarding both the expected magnitude and the expected variability of the deformation to be estimated. We test this model on the problem of left ventricular deformation estimation, and present ideas for its application in image registration and brain deformation during neurosurgery.

Statistical shape models show considerable promise as a basis for segmenting and interpreting images. One of the drawbacks of the approach is, however, the need to establish a set of dense correspondences between examples of similar structures, across a training set of images. Often this is achieved by locating a set of ‘landmarks’ manually on each of the training images, which is time-consuming and subjective for 2D images, and almost impossible for 3D images. This has led to considerable interest in the problem of building a model automatically from a set of training shapes. We extend previous work that has posed this problem as one of optimising a measure of model ‘quality’ with respect to the set of correspondences. We define model ‘quality’ in terms of the information required to code the whole set of training shapes and aim to minimise this description length. We describe a scheme for representing the dense correspondence maps between the training examples and show that a minimum description length model can be obtained by stochastic optimisation. Results are given for several different training sets of 2D boundaries, showing that the automatic method constructs better models than the manual landmarking approach. We also show that the method can be extended straightforwardly to 3D.

This paper presents a Bayesian multi-scale three dimensional deformable template approach based on a medial representation for the segmentation and shape characterization of anatomical objects in medical imagery. Prior information about the geometry and shape of the anatomical objects under study is incorporated via the construction of exemplary templates. The anatomical variability is accommodated in the Bayesian framework by defining probabilistic transformations on these templates. The modeling approach taken in this paper for building exemplary templates and associated transformations is based on a multi-scale medial representation. The transformations defined in this framework are parameterized directly in terms of natural shape operations, such as thickening and bending, and their location. Quantitative validation results are presented on the automatic segmentation procedure developed for the extraction of the kidney parenchyma-including the renal pelvis-in subjects undergoing radiation treatment for cancer. We show that the segmentation procedure developed in this paper is efficient and accurate to within the voxel resolution of the imaging modality.

A novel method is introduced that allows for the generation of landmarks for three-dimensional shapes and the construction of the corresponding 3D Active Shape Models (ASM). Landmarking of a set of examples from a class of shapes is achieved by (i) construction of an atlas of the class, (ii) automatic extraction of the landmarks from the atlas, and (iii) subsequent propagation of these landmarks to each example shape via a volumetric elastic deformation procedure. This paper describes in detail the method to generate the atlas, and the landmark extraction and propagation procedures. This technique presents some advantages over previously published methods: it can treat multiple-part structures, and it requires less restrictive assumptions on the structure’s topology. The applicability of the developed technique is demonstrated with two examples: CT bone data and MR brain data.

A method for regularizing diffusion tensor magnetic resonance images (DT-MRI) is presented. The scheme is divided into two main parts: a restoration of the principal diffusion direction, and a regularization of the 3 eigenvalue maps. The former make use of recent variational methods for restoring direction maps, while the latter makes use of the strong structural information embedded in the diffusion tensor image to drive a non-linear anisotropic diffusion process. The whole process is illustrated on synthetic and real data, and possible improvements are discussed.

A method is presented for determining likely paths of anatomical connection between regions of the brain using MR diffusion tensor information. Level set theory, applied using fast marching methods, is used to generate 3-D time of arrival maps, from which connection paths between brain regions may be identified. The method is demonstrated in the normal brain and it is shown that major white matter tracts may be elucidated and that multiple connections and tract branching are allowed. Maps of the likelihood of connection between brain regions are also determined. Two metrics are described for estimating the (informal) likelihood of connection between regions.

In this paper we propose a novel technique for the analysis of diffusion tensor magnetic resonance images. This method involves solving the full diffusion equation over a finite element mesh derived from the MR data. It calculates connection probabilities between points of interest, which can be compared within or between subjects. Unlike traditional tractography, we use all the data in the diffusion tensor at each voxel which is likely to increase robustness and make intersubject comparisons easier.

A prototype system to assist radiologists in the differential diagnosis of mammographic calcifications is presented. Our approach is to incorporate image-processing operators within a knowledge-based decision support system. The work described in this paper involves three stages. The first is to identify a set of terms that can represent the knowledge required in an example of radiological decision-making. The next is to identify image processing operators to extract the required information from the image. The final stage is to provide links between the set of symbolic terms and the image processing operators.

Neurodegenerative diseases are often associated with loss of brain tissue volume. Our objective was to develop and evaluate a fully automated method to estimate cerebral volume from magnetic resonance images (MRI) of patients with multiple sclerosis (MS). In this study, MRI data from 17 normal subjects and 68 untreated MS patients was used to test the method. Each MRI volume was corrected for image intensity non-uniformity, intensity normalized, brain masked and tissue classified. The classification results were used to compute a normalized metric of cerebral volume based on the Brain to IntraCranial Capacity Ratio (BICCR).This paper shows that the computation of BICCR using automated techniques provides a highly reproducible measurement of relative brain tissue volume that eliminates the need for precise repositioning. Initial results indicate that the measure is both robust and precise enough to monitor MS patients over time to estimate brain atrophy. In addition, brain atrophy may yield a more sensitive endpoint for treatment trials in MS and possibly for other neuro-degenerative diseases such as Huntington’s or Alzheimer’s disease.

The Statistics Based Image Registration (SBR) method for automatic image registration is presented with application to magnetic resonance (MR)and ultrasound (US) cardiac time series images. SBR is demonstrated for MR myocardial perfusion assessment and US myocardial kinetics studies. The utility of the method for a range of other clinical applications is discussed.

This paper presents an extension to the standard Bayesian image analysis paradigm to explicitly incorporate a multiscale approach. This new technique is demonstrated by applying it to the problem of compensating for soft tissue deformation of pre-segmented surfaces for image-guided surgery using 3-D ultrasound. The solution is regularised using knowledge of the mean and Gaussian curvatures of the surface estimate. Results are presented from testing the method on ultrasound data acquired from a volunteer’s liver. Two structures were segmented from an MR scan of the volunteer: the liver surface and the portal vein. Accurate estimates of the deformed surfaces were successfully computed using the algorithm, based on prior probabilities defined using a minimal amount of human intervention. With a more accurate prior model, this technique has the possibility to completely automate the process of compensating for intraoperative deformation in image-guided surgery.

A novel method to obtain correspondence between landmarks when comparing pairs of mammographic images from the same patient is presented. Our approach is based on automatically established correspondence between linear structures (i.e. ducts and vessels) which appear in mammograms using robust features such as orientation, width and curvature extracted from those structures. In addition, a novel multiscale feature matching approach is presented which results in a reliable correspondence between extracted features.

During a neuro-surgical intervention, the brain tissues shift and warp. In order to keep an accurate positioning of the surgical instruments, one has to estimate this deformation from intra-operative images. We present in this article a feasibility study of a tracking tool based on intra-operative 3D ultrasound (US) images. The automatic processing of this kind of images is of great interest for the development of innovative and low-cost image guided surgery tools. The difficulty relies both in the complex nature of the ultrasound image, and in the amount of data to be treated as fast as possible.

Fusing of multi-modal data involves automatically estimating the coordinate transformation required to align the data sets. Most existing methods in literature are not robust and fast enough for practical use. We propose a robust algorithm, based on matching local-frequency image representations, which naturally allow for processing the data at different scales/resolutions, a very desirable property from a computational efficiency view point. This algorithm involves minimizing — over all affine transformations — the integral of the squared error (ISE or L₂E) between a Gaussian model of the residual and its true density function. The residual here refers to the difference between the local frequency representations of the transformed (by an unknown transformation) source and target data. The primary advantage of our algorithm is its ability to cope with large non-overlapping fields of view of the two data sets being registered, a common occurrence in practise. We present implementation results for misalignments between CT and MR brain scans.

Surgical navigation systems provide the surgeon with a display of preoperative and intraoperative data in the same coordinate system. However, the systems currently in use in neurosurgery are subject to inaccuracy caused by intraoperative brain movement (brain shift) since they typically assume that the intracranial structures are rigid. Experiments show brain shift of up to one centimeter, making it the dominant error in the system. We propose a system that compensates for this error. It is based on a continuum 3D biomechanical deformable brain model guided by intraoperative data. The model takes into account neuro-anatomical constraints and is able to correspondingly deform all preoperatively acquired data. The system was tested on two sets of intraoperative MR scans, and an initial validation indicated that our approach reduced the error caused by brain shift.

The present study explores a novel spatiotemporal technique using the fixed effect model for the analysis of functional brain images and propose a novel approach to obtain the least square estimation of the signal subspace of activated voxels. The spatial and temporal domain correlations are incorporated using appropriate prior models and the possibility of using the Markov property to incorporate the spatial domain correlations are investigated.

Spatial and temporal correlations which affect the signal measured in functional MRI (fMRI) are usually not considered simultaneously (i.e., as non-independent random processes) in statistical methods dedicated to detecting cerebral activation.We propose a new method for modeling the covariance of a stationary spatio-temporal random process and apply this approach to fMRI data analysis. For doing so, we introduce a multivariate regression model which takes simultaneously the spatial and temporal correlations into account. We show that an experimental variogram of the regression error process can be fitted to a valid nonseparable spatio-temporal covariance model. This yields a more robust estimation of the intrinsic spatio-temporal covariance of the error process and allows a better modeling of the properties of the random fluctuations affecting the hemodynamic signal. The practical relevance of our model is illustrated using real event-related fMRI experiments.

Dynamic microscopy of the nailfold capillaries using Na- fluorescein dye can be used to assess the condition of the peripheral circulation of Scleroderma patients, yielding more information than simple morphological studies. In this paper we describe a computer based system for this kind of study and present preliminary results on Scleroderma patients. We show how the dye concentrations vary both in time and as a function of distance from the capillary wall in unprecedented resolution, suggesting that a simple permeability model may be applicable to the data.

Clinical magnetic resonance imaging of regional myocardial perfusion has recently become possible with the use of rapid acquisitions to track the kinetics of an intravenous injection of contrast. A great deal of processing is then needed to obtain clinical parameters. In particular, methods to automatically group alike regions for an increased signalto-noise ratio and improved parameter estimates are needed. This work explores two types of time curve analysis techniques for MRI perfusion imaging: factor analysis and clustering. Both methods are shown to work for extraction of the blood input function, with the clustering method appearing to be more robust. The availability of an accurate blood input function then enables more complex approaches to automatically fitting all of the relevant data to appropriate models. These more complex approaches are formulated here and tested in a preliminary fashion.

We present a new approach to detecting functional networks in fMRI time series data. Functional networks as defined here are characterized by a tight coherence criterion where every network member is closely connected to every other member. This definition of a network closely resembles that of a clique in a graph. We propose to use replicator dynamics for detecting such networks. Our approach differs from standard clustering algorithms in that the entities that are targeted here differ from the traditional cluster concept.

Principal component analysis is a well developed and under- stood method of multivariate data processing. Its optimal performance requires knowledge of noise covariance that is not available in most ap- plications. We suggest a method for estimation of noise covariance based on assumed smoothness of the estimated dynamics.

This work provides a new method to estimate and remove baseline drifts in the fMRI signal. The baseline drift in each time series is described as a superposition of physical and physiological phenomena that occur at different scales. A fast algorithm, based on a wavelet representation of the data yields detrended time-series. Experiments with fMRI data demonstrate that our detrending technique can infer and remove drifts that cannot be adequately represented with low degree polynomials. Our detrending technique resulted in a noticeable improvement by reducing the number of false positive and the number of false negative.

Cytoarchitectonic fields of the human neocortex are defined by characteristic variations in the composition of a general six-layer structure. It is commonly accepted that these fields correspond to functionally homogeneous entities. Diligent techniques were developed to characterize cytoarchitectonic fields by staining sections of post-mortem brains and subsequent statistical evaluation. Fields were found to show a considerable interindividual variability in extent and relation to macroscopic anatomical landmarks. With upcoming new high-resolution magnetic resonance (MR) scanning protocols, it appears worthwile to examine the feasibility of characterizing the neocortical fine-structure from anatomical MR scans, thus, defining cytoarchitectonic fields by in-vivo techniques.

This paper describes several experiments that prove that standard motion correction methods may induce spurious activations in some motion-free fMRI studies. This artefact stems from the fact that activated areas behave like biasing outliers for the least square based measure usually driving such registration methods. This effect is demonstrated first using a motion-free simulated time series including artificial activation-like signal changes. Several additional simulations explore the influence of activation on registration accuracy for a wide-range of simulated misregistrations. The effect is finally highlighted on an actual time series obtained from a 3T magnet. All the experiments are performed using four different realignment methods, which allows us to show that the problem is overcome by methods based on robust similarity measures like mutual information.

Event-related functional magnetic resonance imaging (fMRI) is considered as an estimation and reconstruction problem. A linear model of the fMRI system based on the Fourier sampler (k-space) approximation is introduced and used to examine what parameters of the activation are estimable, i.e. can be accurately reconstructed in the noisefree limit. Several possible spatio-temporal representations of the activation are decomposed into null and measurement components. A causal representation of the activation using generalized Laguerre polynomials is introduced.

Little has been done yet to study the synchronization properties of the sources estimated from the MEG/EEG inverse problem, despite the growing interest in the role of phase relations in brain functions. In order to achieve this aim, we propose a novel approach to the MEG/EEG inverse problem based on some regularization using spectral priors: The MEG/EEG raw data are filtered in a frequency band of interest and blurred with some specific “regularization noise” prior to the inversion process. This formalism uses non quadratic regularization and a deterministic optimization algorithm. We proceed to Monte Carlo simulations using the chaotic Rössler oscillators to model the neural generators. Our results demonstate that it is possible to reveal some phase-locking between brain sources with great accuracy following the computation of the inverse problem based on scalp MEG/EEG measurements.

Dynamic systems defined on the scale of neural ensembles are well-suited to model the spatiotemporal dynamics of electroencephalographic (EEG) and magnetoencephalographic (MEG) data. We develop a methodological framework, which defines the activity of neural ensembles, the neural field, on a sphere in three dimensions. Using Magnetic Resonance Imaging (MRI) we map the neural field dynamics from the sphere onto the folded cortical surface of a hemisphere. The neural field represents the current flow perpendicular to the cortex and thus allows the calculation of the electric potentials on the surface of the skull and the magnetic fields outside the skull to be measured by EEG and MEG, respectively. For demonstration of the dynamics, we present the propagation of activation at a single cortical site resulting from a transient input. Non-trivial mappings between the multiple levels of observation are obtained which would not be predicted by inverse solution techniques. Considering recent results mapping large-scale brain dynamics (EEG, MEG) onto behavioral motor patterns, this paper provides a discussion of the causal chain starting from local neural ensemble dynamics through encephalographic data to behavior.

This paper describes the design, implementation and preliminary results of a unified non-rigid feature registration method for the purpose of brain anatomical structure alignment. We combine different types of features together and fuse them into a common point representation. This enables the co-registration of all features using a new non-rigid point matching algorithm. In this way, the spatial interrelationships between different features are directly utilized to improve the registration accuracy. We also conducted a carefully designed synthetic study to compare some anatomical features’ ability for non-rigid brain structure alignment. This study allows us to evaluate the relative improvements in registration accuracy when different features are combined.

In this paper, we investigate the introduction of cortical constraints for non rigid inter-subject brain registration. We extract sulcal patterns with the active ribbon method, presented in [10]. An energy based registration method [7] makes it possible to incorporate the matching of cortical sulci, and express in a unified framework the local sparse similarity and the global “iconic” similarity. We show the objective benefits of cortical constraints on a database of 18 subjects, with global and local measures of the quality of the registration.

Landmark-based thin-plate spline image registration is one of the most commonly used methods for non-rigid medical image registration and anatomical shape analysis. It is well known that this method does not produce a unique correspondence between two images away from the landmark locations because interchanging the role of source and target landmarks does not produce forward and reverse transformations that are inverses of each other. In this paper, we present two new image registration algorithms that minimize the thin-plate spline bending energy and the inverse consistency error—the error between the forward and the inverse of the reverse transformation. The landmarkbased consistent thin-plate spline algorithm registers images given a set of corresponding landmarks while the intensity-based consistent thinplate spline algorithm uses both corresponding landmarks and image intensities. Results are presented that demonstrate that using landmark and intensity information to jointly estimate the forward and reverse transformations provides better correspondence than using landmarks or intensity alone.

We present a novel validation method for non-rigid registration using a simulation of deformations based on biomechanical modelling of tissue properties. This method is tested on a previously developed non-rigid registration method for dynamic contrast enhanced Magnetic Resonance (MR) mammography image pairs [1]. We have constructed finite element breast models and applied a range of displacements to them, with an emphasis on generating physically plausible deformations which may occur during normal patient scanning procedures. From the finite element method (FEM) solutions, we have generated a set of deformed contrast enhanced images against which we have registered the original dynamic image pairs. The registration results have been successfully validated at all breast tissue locations by comparing the recovered displacements with the biomechanical displacements. The validation method presented in this paper is an important tool to provide biomechanical gold standard deformations for registration error quantification, which may also form the basis to improve and compare different non-rigid registration techniques for a diversity of medical applications.

A sequential algorithm is presented for computing the Euclidean distance transform of a k-dimensional binary image in time linear in the total number of voxels. The algorithm may be of practical value since it is relatively simple and easy to implement and it is relatively fast (not only does it run in linear time but the time constant is small).

The geometric deformable model (GDM) provides a useful framework for segmentation by integrating the energy minimization concept of classical snakes with the topologically flexible gradient flow. The key aspect of this technique is the image derived conformal metric for the configuration space. While the theoretical and numerical aspects of the geometric deformable model have been discussed in the literature, the formation of the conformal metric itself has not received much attention. Previous definitions of the conformal metric do not allow the GDM to produce reliable segmentation results in low-contrast or highblur regions. This paper examines the desired properties of the conformal metric with regard to the image information and proposes an elliptic partial differential equation to construct the metric. Our method produces similar results to other metric definitions in high-contrast regions, but produces better results in low-contrast, high-blur situations.

Supplied with image data from two distinct populations we apply a non-rigid registration technique to place each image into correspondence with an atlas. Having found the appropriate transformations we then use the use determinant of the Jacobian of the corresponding transformations and find the linear discriminant function which can best distinguish between the populations on the basis of this data. We apply the method to a collection of mid-sagittal slices of the corpus callosum for a group of 34 males and 52 females. We find that there appear to be no statistically significant differences between the relative sizes of regions in the corpus callosum between males and females.

To improve the robustness of segmentation methods, more and more methods use prior knowledge. We present an approach which embeds an active shape model into an elastically deformable surface model, and combines the advantages of both approaches. The shape model constrains the flexibility of the surface mesh representing the deformable model and maintains an optimal distribution of mesh vertices. A specific external energy which attracts the deformable model to locally detected surfaces, reduces the danger that the mesh is trapped by false object boundaries. Examples are shown, and furthermore a validation study for the segmentation of vertebrae in CT images is presented. With the exception of a few problematic areas, the algorithm leads reliably to a very good overall segmentation.

The prevalent model for second order variation in 3-D volumes is an ellipsoid spanned by the magnitudes of the Hessian eigenvalues. Here, we describe this variation as a vector in an orthogonal shape space spanned by spherical harmonic basis functions. From this newsh ape-space, a truly rotation- and shape-invariant signal energy is defined, consistent orientation information is extracted and shape sensitive quantities are employed. The advantage of these quantities is demonstrated in detection of stenosis in Magnetic Resonance Angiography( MRA) volume. The news hape space is expected to improve both the theoretical understanding and the implementation of Hessian based analysis in other applications as well.

Reconstructing a topologically correct representation of the brain cortex surface from magnetic resonance images is important in several medical and neuroscience applications. Most previous methods have either made drastic changes to the underlying anatomy or relied on hand-editing. Recently, a new technique due to Shattuck and Leahy yields a fully-automatic procedure with little distortion of the underlying segmentation. The present paper can be considered as an extension of this approach to include arbitrary cut directions and arbitrary digital connectivities. A detailed analysis of the method’s performance on 15 magnetic resonance brain images is provided.

Analysis of shape variability is important for diagnostic classification and understanding of biological processes. We present a novel shape analysis approach based on a multiscale medial representation. Our method examines shape variability in separate categories, such as global variability in the coarse-scale shape description and localized variability in the fine-scale description. The method can distinguish between variability in growing and bending. When used for diagnostic classification, the method indicates what shape change accounts for the discrimination and where on the object the change occurs. We illustrate the approach by analysis of 2D clinical corpus callosum shape and discrimination of simulated corpora callosa.

Skeleton is a frequently applied shape feature to represent the general form of an object. Thinning is an iterative object reduction technique for producing a reasonable approximation to the skeleton in a topology preserving way. This paper describes a sequential 3D thinning algorithm for extracting medial lines of objects in (26, 6) pictures. Our algorithm has been successfully applied in medical image analysis. Three of the emerged applications (analysing airways, blood vessels, and colons) are also presented.

An efficient adaptive multigrid level set method for front propagation purposes in three dimensional medical image segmentation is presented. It is able to deal with non sharp segment boundaries. A flexible, interactive modulation of the front speed depending on various boundary and regularization criteria ensure this goal. Efficiency is due to a graded underlying mesh implicitly defined via error or feature indicators. A suitable saturation condition ensures an important regularity condition on the resulting adaptive grid. As a casy study the segmentation of glioma is considered. The clinician interactively selects a few parameters describing the speed function and a few seed points. The automatic process of front propagation then generates a family of segments corresponding to the evolution of the front in time, from which the clinician finally selects an appropriate segment covered by the gliom. Thus, the overall glioma segmentation turns into an efficient, nearly real time process with intuitive and usefully restricted user interaction.

A mixture model clustering algorithm is presented for robust MRI brain image segmentation in the presence of partial volume averaging. The method uses additional classes to represent partial volume voxels of mixed tissue type in the data with their probability distributions modeled accordingly. The image model also allows for tissue-dependent variance values and voxel neighborhood information is taken into account in the clustering formulation. The final result is the estimated fractional amount of each tissue type present within a voxel in addition to the label assigned to the voxel. A multi-threaded implementation of the method is evaluated using both synthetic and real MRI data.

An approach is described which has the potential to unify edge preserving smoothing with segmentation based on differential edge detection at multiple scales. The analysis of n-D data is decomposed into independent 1-D problems. Smoothing in various directions along 1-D profiles through n-D data is driven by local structure separation, rather than by local contrast. Analytic expressions are obtained for the derivatives of the edge preserved 1-D profiles. Using these expressions, multidimensional edge detection operators such as the Laplacian or second directional derivative can be composed and used to segment n-D data. The smoothing and segmentation algorithms are applied to simulated 4-D medical images.

This paper presents a new approach for the automatic segmentation and characterization of active MS lesions in 4D data of multiple sequences. Traditional segmentation of 4D data applies individual 3D spatial segmentation to each image data set, thus not making use of correlation over time. More recently, a time series analysis has been applied to 4D data to reveal active lesions [3]. However, misregistration at tissue borders led to false positive lesion voxels.Lesion development is a complex spatio-temporal process, consequently methods concentrating exclusively on the spatial or temporal aspects of it cannot be expected to provide optimal results. Active MS lesions 44 were extracted from the 4D data in order to quantify MR-based spatiotemporal changes in the brain. A spatio-temporal lesion model generated by principal component analysis allowed robust identification of active MS lesions overcoming the drawbacks of traditional purely spatial or purely temporal segmentation methods.

This paper describes a novel, 2D+time Active Appearance Motion Model (AAMM). Cootes’s 2D AAM framework was extended by considering a complete image sequence as a single shape/intensity sample. This way, the cardiac motion is modeled in combination with the shape and image appearance of the heart. The clinical potential of the AAMMs is demonstrated on two imaging modalities — cardiac MRI and echocardiography.

We propose a novel feature enhancement approach to enhance the quality of noisy images. It is based on a phase-based feature detection algorithm, followed by sparse surface interpolation and subsequent nonlinear post-processing. We first exploit the intensity-invariant property of phase-based acoustic feature detection to select a set of relevant image features in the data. Then, an approximation to the low frequency components of the sparse set of selected features is obtained using a fast surface interpolation algorithm. Finally, a non-linear postprocessing step is applied. Results of applying the method to echocardiographic sequences (2D+T) are presented. We show that the correction is consistent over time and does not introduce any artefacts. An evaluation protocol is proposed in the case of echocardiographic data and quantitative results are presented.

This paper presents a new and integrated approach to automatic 3D brain vessel segmentation using physics-based statistical models of background and vascular signals, and velocity (flow) field information in phase contrast magnetic resonance angiograms (PC-MRA). The proposed new approach makes use of realistic statistical models to detect vessels more accurately than conventional intensity gradient-based approaches. In this paper, rather than using MRA speed images alone, as in prior work [7,8,10], we define a 3D local phase coherence (LPC) measure to incorporate velocity field information. The proposed new approach is an extension of our previous work in 2D vascular segmentation [5,6], and is formulated in a variational framework, which is implemented using the recently proposed modified level set method [1]. Experiments on flow phantoms, as well as on clinical data sets, show that our approach can segment normal vasculature as well as low flow (low SNR) or complex flow regions, especially in an aneurysm.

This paper investigates the segmentation of different regions in PET images based on the feature vector extracted from the timeactivity curve for each voxel. PET image segmentation has applications in PET reference region analysis and activation studies. The segmentation algorithm presented uses a Markov random field model for the voxel class labels. By including the Markov random field model in the expectation-maximisation iteration, the algorithm can be used to simultaneously estimate parameters and segment the image. Hence, the algorithm is able to combine both feature and spatial information for the purpose of segmentation. Experimental results on synthetic and real PET data are presented to demonstrate the performance of the algorithm. The algorithms used in this paper can be used to segment other functional images.

A method for building a statistical shape model of sulci of the human brain cortex is described. The model includes sulcal fundi that are defined on a spherical map of the cortex. The sulcal fundi are first extracted in a semi-automatic way using an extension of the fast marching method. They are then transformed to curves on the unit sphere via a conformal mapping method that maps each cortical point to a point on the unit sphere. The curves that represent sulcal fundi are parameterized with piecewise constant-speed parameterizations. Intermediate points on these curves correspond to sulcal landmarks, which are used to build a point distribution model on the unit sphere. Statistical information of local properties of the sulci, such as curvature and depth, are embedded in the model. Experimental results are presented to show how the models are built.

The rapid creation of comprehensive brain image databases mandates the development of mathematical algorithms to uncover diseasespecific patterns of brain structure and function in human populations. We describe our construction of probabilistic atlases that store detailed information on how the brain varies across age and gender, across time, in health and disease, and in large human populations. Specifically, we introduce a mathematical framework based on covariant partial differential equations (PDEs), pull-backs of mappings under harmonic flows, and high-dimensional random tensor fields to encode variations in cortical patterning, asymmetry and tissue distribution in a population-based brain image database (N=94 scans). We use this information to detect disease-specific abnormalities in Alzheimer’s disease and schizophrenia, including dynamic changes over time. Illustrative examples are chosen to show how group patterns of cortical organization, asymmetry, and disease-specific trends can be resolved that are not apparent in individual brain images. Finally, we create four-dimensional (4D) maps that store probabilistic information on the dynamics of brain change in development and disease. Digital atlases that generate these maps show considerable promise in identifying general patterns of structural and functional variation in diseased populations, and revealing how these features depend on demographic, genetic, clinical and therapeutic parameters.

Knowledge about the biological variability of anatomical objects is essential for statistical shape analysis and discrimination between healthy and pathological structures. This paper describes a novel approach that incorporates variability of an object population into the generation of a characteristic 3D shape model. The proposed shape representation is based on a fine-scale spherical harmonics (SPHARM) boundary description and a coarse-scale sampled medial description. The medial description is composed of a net of medial samples (m-rep) with fixed graph properties. The medial model is computed automatically from a predefined shape space using pruned 3D Voronoi skeletons to determine the stable medial branching topology. An intrinsic coordinate system and an implicit correspondence between shapes is defined on the medial manifold. Our novel representation describes shape and shape changes in a natural and intuitive fashion. Several experimental studies of biological structures regarding shape asymmetry and similarity clearly demonstrate the meaningful represesentation of local and global form.

Statistical analysis of anatomical shape differences between two different populations can be reduced to a classification problem, i.e., learning a classifier function for assigning new examples to one of the two groups while making as few mistakes as possible. In this framework, feature vectors representing the shape of the organ are extracted from the input images and are passed to the learning algorithm. The resulting classifier then has to be interpreted in terms of shape differences between the two groups back in the image domain. We propose and demonstrate a general approach for such interpretation using deformations of outline meshes to represent shape differences. Given a classifier function in the feature space, we derive a deformation that corresponds to the differences between the two classes while ignoring shape variability within each class. The algorithm essentially estimates the gradient of the classification function with respect to node displacements in the outline mesh and constructs the deformation of the mesh that corresponds to moving along the gradient vector. The advantages of the presented algorithm include its generality (we derive it for a wide class of non-linear classifiers) as well as its flexibility in the choice of shape features used for classification. It provides a link from the classifier in the feature space back to the natural representation of the original shapes as surface meshes. We demonstrate the algorithm on artificial examples, as well as a real data set of the hippocampus-amygdala complex in schizophrenia patients and normal controls.