Medical Image Computing and Computer-Assisted Intervention – MICCAI 2012

Model-based segmentation approaches have been proven to produce very accurate segmentation results while simultaneously providing an anatomic labeling for the segmented structures. However, variations of the anatomy, as they are often encountered e.g. on the drainage pattern of the pulmonary veins to the left atrium, cannot be represented by a single model. Automatic model selection extends the model-based segmentation approach to handling significant variational anatomies without user interaction. Using models for the three most common anatomical variations of the left atrium, we propose a method that uses an estimation of the local fit of different models to select the best fitting model automatically. Our approach employs the support vector machine for the automatic model selection. The method was evaluated on 42 very accurate segmentations of MRI scans using three different models. The correct model was chosen in 88.1 % of the cases. In a second experiment, reflecting average segmentation results, the model corresponding to the clinical classification was automatically found in 78.0 % of the cases.

Curvilinear structures are common in medical imaging, which typically require dedicated processing techniques. We present a new structure to process these, that we call the polygonal path image, denoted

$\mathfrak{P}$

. We derive from

$\mathfrak{P}$

some curvilinear structure enhancement and analysis algorithms. We show that

$\mathfrak{P}$

has some interesting properties: it generalizes several concepts found in other methods; it makes it possible to control the smoothness and length of the structures under study; and it can be computed efficiently. We estimate quantitatively its performance in the context of interventional cardiology for the detection of guide-wires in X-ray images. We show that

$\mathfrak{P}$

is particularly well suited for this task where it appears to outperform previous state of the art techniques.

Dynamic overlay of 3D models onto 2D X-ray images has important applications in image guided interventions. In this paper, we present a novel catheter tracking for motion compensation in the Transcatheter Aortic Valve Implantation(TAVI). To address such challenges as catheter shape and appearance changes, occlusions, and distractions from cluttered backgrounds, we present an adaptive linear discriminant learning method to build a measurement model online to distinguish catheters from background. An analytic solution is developed to effectively and efficiently update the discriminant model and to minimize the classification errors between the tracking object and backgrounds. The online learned discriminant model is further combined with an offline learned detector and robust template matching in a Bayesian tracking framework. Quantitative evaluations demonstrate the advantages of this method over current state-of-the-art tracking methods in tracking catheters for clinical applications.

Minimally invasive cardiac surgery is made possible by image guidance technology. X-ray fluoroscopy provides high contrast images of catheters and devices, whereas 3D ultrasound is better for visualising cardiac anatomy. We present a system in which the two modalities are combined, with a trans-esophageal echo volume registered to and overlaid on an X-ray projection image in real-time. We evaluate the accuracy of the system in terms of both temporal synchronisation errors and overlay registration errors. The temporal synchronisation error was found to be 10% of the typical cardiac cycle length. In 11 clinical data sets, we found an average alignment error of 2.9 mm. We conclude that the accuracy result is very encouraging and sufficient for guiding many types of cardiac interventions. The combined information is clinically useful for placing the echo image in a familiar coordinate system and for more easily identifying catheters in the echo volume.

Current treatments of heart rhythm troubles require careful planning and guidance for optimal outcomes. Computational models of cardiac electrophysiology are being proposed for therapy planning but current approaches are either too simplified or too computationally intensive for patient-specific simulations in clinical practice. This paper presents a novel approach, LBM-EP, to solve any type of mono-domain cardiac electrophysiology models at near real-time that is especially tailored for patient-specific simulations. The domain is discretized on a Cartesian grid with a level-set representation of patient’s heart geometry, previously estimated from images automatically. The cell model is calculated node-wise, while the transmembrane potential is diffused using Lattice-Boltzmann method within the domain defined by the level-set. Experiments on synthetic cases, on a data set from CESC’10 and on one patient with myocardium scar showed that LBM-EP provides results comparable to an FEM implementation, while being 10 − 45 times faster. Fast, accurate, scalable and requiring no specific meshing, LBM-EP paves the way to efficient and detailed models of cardiac electrophysiology for therapy planning.

Patient-specific cardiac modelling can help in understanding pathophysiology and predict therapy planning. However it requires to personalize the model geometry, kinematics, electrophysiology and mechanics. Calibration aims at providing global values (space invariant) of parameters before performing the personalization stage which involves solving an inverse problem to find regional values. We propose an automatic calibration method of the mechanical parameters of the Bestel-Clément-Sorine (BCS) electromechanical model of the heart based on the Unscented Transform algorithm. A sensitivity analysis is performed that reveals which observations on the volume and pressure evolution are significant to characterize the global behaviour of the myocardium. We show that the calibration method gives satisfying results by optimizing up to 7 parameters of the BCS model in only one iteration. This method was evaluated on 7 volunteers and 2 heart failure patients, with a mean relative error from the real data of 11%. This calibration enabled furthermore a preliminary study of the specific parameters to the studied pathologies.

In this paper, we adapt spectral signatures for capturing morphological changes over time. Advanced techniques for capturing temporal shape changes frequently rely on first registering the sequence of shapes and then analyzing the corresponding set of high dimensional deformation maps. Instead, we propose a simple encoding motivated by the observation that small shape deformations lead to minor refinements in the spectral signature composed of the eigenvalues of the Laplace operator. The proposed encoding does not require registration, since spectral signatures are invariant to pose changes. We apply our representation to the shapes of the ventricles extracted from 22 cine MR scans of healthy controls and Tetralogy of Fallot patients. We then measure the accuracy score of our encoding by training a linear classifier, which outperforms the same classifier based on volumetric measurements.

We present an extension of the diffeomorphic Geometric Demons algorithm which combines the iconic registration with geometric constraints. Our algorithm works in the log-domain space, so that one can efficiently compute the deformation field of the geometry. We represent the shape of objects of interest in the space of currents which is sensitive to both location and geometric structure of objects. Currents provides a distance between geometric structures that can be defined without specifying explicit point-to-point correspondences. We demonstrate this framework by registering simultaneously

T

1

images and 65 fiber bundles consistently extracted in 12 subjects and compare it against non-linear

T

1

, tensor, and multi-modal

T

1

 + Fractional Anisotropy (FA) registration algorithms. Results show the superiority of the Log-domain Geometric Demons over their purely iconic counterparts.

This paper presents a novel skeleton based method for the registration of head&neck datasets. Unlike existing approaches it is fully automated, spatial relation of the bones is considered during their registration and only one of the images must be a CT scan. An articulated atlas is used to jointly obtain a segmentation of the skull, the mandible and the vertebrae C1-Th2 from the CT image. These bones are then successively rigidly registered with the moving image, beginning at the skull, resulting in a rigid transformation for each of the bones. Linear combinations of those transformations describe the deformation in the soft tissue. The weights for the transformations are given by the solution of the Laplace equation. Optionally, the skin surface can be incorporated. The approach is evaluated on 20 CT/MRI pairs of head&neck datasets acquired in clinical routine. Visual inspection shows that the segmentation of the bones was successful in all cases and their successive alignment was successful in 19 cases. Based on manual segmentations of lymph nodes in both modalities, the registration accuracy in the soft tissue was assessed. The mean target registration error of the lymph node centroids was 5.33 ± 2.44 mm when the registration was solely based on the deformation of the skeleton and 5.00 ± 2.38 mm when the skin surface was additionally considered. The method’s capture range is sufficient to cope with strongly deformed images and it can be modified to support other parts of the body. The overall registration process typically takes less than 2 minutes.

Accurate registration of human lungs in CT images is required for many applications in pulmonary image analysis and used for example for atlas generation. While various registration approaches have been developed in the past, the correct alignment of the interlobular fissures is still challenging for many reasons, especially for inter-patient registration. Fissures are depicted with very low contrast and their proximity in the image shows little detail due to the lack of vessels. Moreover, iterative registration algorithms usually require the objects to be overlapping in both images to find the right transformation, which is often not the case for fissures.

In this work, a novel approach is presented for integrated lobe segmentation and intensity-based registration aiming for a better alignment of the interlobular fissures. To this end, level sets with a shape-based fissure attraction term are used to formulate a new condition in the registration framework. The method is tested for pairwise registration of lung CT scans of nine different subjects and the results show a significantly improved matching of the pulmonary lobes after registration.

Accurate real-time registration of intra-operative ultrasound (US) to computed tomography (CT) remains a challenging problem. In orthopedic applications, a recent promising approach proposed the use of Gaussian mixture modeling for bone surface registration. Though relatively successful, the method relied on naïve and error prone subsampling of the surfaces registered to reduce computational cost and also heavily relied on heuristically-set parameters for bone surface generation. In this paper, we present an improved approach employing a novel point simplification method that redistributes surface points to better represent the surface achieving near real-time registration with higher accuracy and robustness. We also present a framework for automating the parameter selection in the bone surface extraction step. For validation, we present extensive quantitative tests on phantom and clinical data obtained by scanning patients with pelvic ring fractures in the operating room. We show an 89% average improvement in target registration error over the recent GMM registration based method.

Deformable registration has been widely used in neuroscience studies for spatial normalization of brain images onto a standard space. Due to high anatomical variances across individual brains, registration performance could be limited when trying to estimate entire deformation pathway either from template to subject or subject to template. Symmetric image registration offers an effective way to simultaneously deform template and subject images towards each other until they meet at the middle point. Although some intensity-based registration algorithms have nicely incorporated this concept of symmetric deformation, the intensity matching between two images may not necessarily imply the correct matching of anatomical correspondences. In this paper, we integrate both strategies of hierarchical attribute matching and symmetric diffeomorphic deformation for building a new symmetric-diffeomorphic registration algorithm for MR brain images. The performance of our proposed method has been extensively evaluated and further compared with top-ranked image registration methods (SyN and diffeomorphic Demons) on brain MR images. In all experiments, our registration method achieves the best registration performance, compared to all other state-of-the-art registration methods.

We formulate the pigmented-skin-lesion (PSL) matching problem as a relaxed labeling of an association graph. In this graph labeling problem, each node represents a mapping between a PSL from one image to a PSL in the second image and the optimal labels are those optimizing a high order Markov Random Field energy (MRF). The energy is made up of unary, binary, and ternary energy terms capturing the likelihood of matching between the points, edges, and cliques of two graphs representing the spatial distribution of the two PSL sets. Following an exploration of various MRF energy terms, we propose a novel entropy energy term encouraging solutions with low uncertainty. By interpreting the relaxed labeling as a measure of confidence, we further leverage the high confidence matching to sequentially constrain the learnt objective function defined on the association graph. We evaluate our method on a large set of synthetic data as well as 56 pairs of real dermatological images. Our proposed method compares favorably with the state-of-the-art.

A method for categorizing landmark-local appearances extracted from computed tomography (CT) datasets is presented. Anatomical landmarks in the human body inevitably have inter-individual variations that cause difficulty in automatic landmark detection processes. The goal of this study is to categorize subjects (i.e., training datasets) according to local shape variations of such a landmark so that each subgroup has less shape variation and thus the machine learning of each landmark detector is much easier. The similarity between each subject pair is measured based on the non-rigid registration result between them. These similarities are used by the spectral clustering process. After the clustering, all training datasets in each cluster, as well as synthesized intermediate images calculated from all subject-pairs in the cluster, are used to train the corresponding subgroup detector. All of these trained detectors compose a detector ensemble to detect the target landmark. Evaluation with clinical CT datasets showed great improvement in the detection performance.

A new approach to align 3D CT data of a segmented lung object with a given prototype (reference lung object) using an affine transformation is proposed. Visual appearance of the lung from CT images, after equalizing their signals, is modeled with a new 3D Markov-Gibbs random field (MGRF) with pairwise interaction model. Similarity to the prototype is measured by a Gibbs energy of signal co-occurrences in a characteristic subset of voxel pairs derived automatically from the prototype. An object is aligned by an affine transformation maximizing the similarity by using an automatic initialization followed by a gradient search. Experiments confirm that our approach aligns complex objects better than popular conventional algorithms.

Image registration is inherently ill-posed, and lacks a unique solution. In the context of medical applications, it is desirable to avoid solutions that describe physically unsound deformations within the patient anatomy. Among the accepted methods of regularizing non-rigid image registration to provide solutions applicable to medical practice is the penalty of thin-plate bending energy. In this paper, we develop an exact, analytic method for computing the bending energy of a three-dimensional B-spline deformation field as a quadratic matrix operation on the spline coefficient values. Results presented on ten thoracic case studies indicate the analytic solution is between 61–1371x faster than a numerical central differencing solution.

Locally affine (polyaffine) image registration methods capture intersubject non-linear deformations with a low number of parameters, while providing an intuitive interpretation for clinicians. Considering the mandible bone, anatomical shape differences can be found at different scales, e.g. left or right side, teeth, etc. Classically, sequential coarse to fine registration are used to handle multiscale deformations, instead we propose a simultaneous optimization of all scales. To avoid local minima we incorporate a prior on the polyaffine transformations. This kind of groupwise registration approach is natural in a polyaffine context, if we assume one configuration of regions that describes an entire group of images, with varying transformations for each region. In this paper, we reformulate polyaffine deformations in a generative statistical model, which enables us to incorporate deformation statistics as a prior in a Bayesian setting. We find optimal transformations by optimizing the maximum a posteriori probability. We assume that the polyaffine transformations follow a normal distribution with mean and concentration matrix. Parameters of the prior are estimated from an initial coarse to fine registration. Knowing the region structure, we develop a blockwise pseudoinverse to obtain the concentration matrix. To our knowledge, we are the first to introduce simultaneous multiscale optimization through groupwise polyaffine registration. We show results on 42 mandible CT images.

Diffusion tensor imaging is widely used in brain connectivity study. As more and more group studies recruit a large number of subjects, it is important to design registration methods that are not only theoretically rigorous, but also computationally efficient, for processing large data sets. However, the requirement of reorienting diffusion tensors complicates and slows down the registration, especially for those methods whose scalar-image versions have linear complexity, for example, the Demons algorithm. In this paper, we propose an extension of the Demons algorithm that incorporates exact reorientation and regularization into the calculation of deforming velocity, yet preserving its linear complexity. This method restores the computational efficiency of the Demons algorithm to diffusion images, but does not sacrifice registration goodness. In our experiments, the new algorithm achieved state-of-art performance at a ten-fold decrease of computational time.

Adolescent Idiopathic Scoliosis (AIS) characterized by the 3D spine deformity affects about 4% schoolchildren worldwide. Several studies have demonstrated the malfunctioning of postural balance, proprioception, and equilibrium control in patients with AIS. Since these functions are closely related to structures in and around the brainstem, the morphometry of the brainstem surface is of utmost importance. In this paper, we propose an effective method to accurately compute the registration between brainstem surfaces. Four consistent features, which describe the global geometry of the brainstem, are automatically extracted to guide the surface registration. Using the discrete Ricci flow method, brainstem surfaces are parameterized conformally onto the quadrilaterally-faced hexahedron, through which the brainstem registration can be obtained. Our registration algorithm can guarantee the exact landmark correspondence between brainstem surfaces. With the obtained registration, a shape energy can be defined to measure the local shape difference between different brainstem surfaces. We have tested our algorithms on 30 real brainstem surfaces extracted from MRIs of 15 normal subjects and 15 AIS patients. Experimental results show the efficacy of the proposed algorithm to register brainstem surfaces, which matches landmark features consistently. The computed registration can be used for the morphometry of brainstems.

We present an extension of the symmetric ICP algorithm that is unbiased for an arbitrary number (

N

 ≥ 2) of shapes, using rigid transformations and scaling. The method does not require the selection of a reference shape or registration order and hence it is unbiased towards any of the registered shapes. The functional to be minimized is non-linear in the transformation parameters and thus computationally complex. We therefore propose a first order approximation that estimates the transformation parameters in a closed form, with computational complexity

$\mathcal{O}(N^{2})$

.

Using a set of wrist bones, we show that the least-squares minimization and the proposed approximation converge to the same solution. Experiments also show that the proposed algorithms lead to smaller registration errors than algorithms that select a reference shape or register to an evolving mean shape. The low computational cost and trivial parallelization enable the alignment of large numbers of bones.

We propose an iterative two-step method to compute a diffeomorphic non-rigid transformation between images of anatomical structures with rigid parts, without any user intervention or prior knowledge on the image intensities. First we compute spatially sparse, locally optimal rigid transformations between the two images using a new block matching strategy and an efficient numerical optimiser (BOBYQA). Then we derive a dense, regularised velocity field based on these local transformations using matrix logarithms and M-smoothing. These two steps are iterated until convergence and the final diffeomorphic transformation is defined as the exponential of the accumulated velocity field. We show our algorithm to outperform the state-of-the-art log-domain diffeomorphic demons method on dynamic cervical MRI data.

Registration of Diffusion-weighted imaging (DWI) data emerges as an important topic in magnetic resonance (MR) image analysis. As existing methods are often designed for specific diffusion models, it is difficult to fit to the registered data different models other than the one used for registration. In this paper we describe a diffeomorphic registration algorithm for DWI data in a large deformation setting. Our method generates spatially normalized DWI data and it is thus possible to fit various diffusion models

after

registration for comparison purposes. Our algorithm includes (1) a reorientation component, where each diffusion profile (DWI signal as a function on a unit sphere) is decomposed, reoriented and recomposed to form the orientation-corrected DWI profile, and (2) a large deformation diffeomorphic registration component to ensure one-to-one mapping in a large-structural-variation scenario. In addition our algorithm uses a geodesic shooting mechanism to avoid the huge computational resources that are needed to register high-dimensional vector-valued data. We also incorporate into our algorithm a multi-kernel strategy where anatomical structures at different scales are considered simultaneously during registration. We demonstrate the efficacy of our method using

in vivo

data.

We present a new method for the estimation of the next brain MR scan in a longitudinal tumor follow-up study. Our method effectively incorporates information of the past scans in the time series to predict the future scan of the patient. Its advantages are that it requires no user intervention and does not assume any particular tumor growth model. Instead, the patient-specific tumor growth parameters are estimated individually from the past patient scans. To validate our method, we conducted an experimental study on four patients with Optic Path Gliomas (OPGs) and four patients with glioblastomas multiforma (GBM), each scanned at five time points. The tumor volumes in the predicted and actual future scans, both segmented by an expert radiologist, yield a mean volume overlap difference of 13.65% for the OPG patients, and 34.23% for the GBM patients.

We propose a new method to parcellate the cerebral cortex based on spatial dependancy in the fluctuations observed with functional Magnetic Resonance Imaging (fMRI) during rest. Our surface-based approach uses a region growing method. In contrast to previous methods, locally stable seed points are identified on the cortical surface and these are grown into a (relatively large 1000 to 5000) number of spatially contiguous regions on both hemispheres. Spatially constrained hierarchical clustering is then used to further combine these regions in a hierarchical tree. Using short-TR resting state fMRI data, this approach allows a

subject specific

parcellation of the cortex into anatomically plausible subregions, identified with high scan-to-scan reproducibility and with borders that delineate clear changes in functional connectivity.

We propose a framework for quantifying node-level community structures between groups using anatomical brain networks derived from DTI-tractography. To construct communities, we computed hierarchical binary trees by maximizing two metrics: the well-known modularity metric (Q), and a novel metric that measures the difference between inter-community and intra-community path lengths. Changes in community structures on the nodal level were assessed between generated trees and a statistical framework was developed to detect local differences between two groups of community structures. We applied this framework to a sample of 42 subjects with major depression and 47 healthy controls. Results showed that several nodes (including the bilateral precuneus, which have been linked to self-awareness) within the default mode network exhibited significant differences between groups. These findings are consistent with those reported in previous literature, suggesting a higher degree of ruminative self-reflections in depression.

In this paper, anatomical development is modeled as a collection of distinctive image patterns localized in space and time. A Bayesian posterior probability is defined over a random variable of subject age, conditioned on data in the form of scale-invariant image features. The model is automatically learned from a large set of images exhibiting significant variation, used to discover anatomical structure related to age and development, and fit to new images to predict age. The model is applied to a set of 230 infant structural MRIs of 92 subjects acquired at multiple sites over an age range of 8-590 days. Experiments demonstrate that the model can be used to identify age-related anatomical structure, and to predict the age of new subjects with an average error of 72 days.

Mild cognitive impairment (MCI) is difficult to diagnose due to its subtlety. Recent emergence of advanced network analysis techniques utilizing resting-state functional Magnetic Resonance Imaging (rs-fMRI) has made the understanding of neurological disorders more comprehensively at a whole-brain connectivity level. However, inferring effective brain connectivity from fMRI data is a challenging task, particularly when the ultimate goal is to obtain good control-patient classification performance. Incorporating sparsity into connectivity modeling can potentially produce results that are biologically more meaningful since most biologically networks are formed by a relatively few number of connections. However, this constraint, when applied at an individual level, will degrade classification performance due to inter-subject variability. To address this problem, we consider a constrained sparse linear regression model associated with the least absolute shrinkage and selection operator (LASSO). Specifically, we introduced sparsity into brain connectivity via

l

1

-norm penalization, and ensured consistent non-zero connections across subjects via

l

2

-norm penalization. Our results demonstrate that the constrained sparse network gives better classification performance than the conventional correlation-based network, indicating its greater sensitivity to early stage brain pathologies.

For clinical evaluation, assessing amyloid deposition with PiB-PET is desirable without requiring MR acquisition and associated fusion/segmentation techniques. A useful clinical tool is to estimate PiB-PET against the brain surface, which is however challenging using PET alone because of the lack of structural information. We propose a method to generate such estimate, where multiple atlases are selected and combined with local weights in a Bayesian framework. Qualitative and quantitative comparison with and without MRI are presented. Using PET only, the average error on the brain surface was around 13% compared to MRI-dependant method.

In this study, we propose a framework to map functional MRI (fMRI) activation signals using DTI-tractography. This framework, which we term functional by structural hierarchical (FSH) mapping, models the regional origin of fMRI brain activation to construct “N-step reachable structural maps”. Linear combinations of these N-step reachable maps are then used to predict the observed fMRI signals. Additionally, we constructed a utilization matrix, which numerically estimates whether the inclusion of a specific structural connection better predicts fMRI, using simulated annealing. We applied this framework to a visual fMRI task in a sample of body dysmorphic disorder (BDD) subjects and comparable healthy controls. Group differences were inferred by comparing the observed utilization differences against 10,000 permutations under the null hypothesis. Results revealed that BDD subjects under-utilized several key local connections in the visual system, which may help explain previously reported fMRI findings and further elucidate the underlying pathophysiology of BDD.

Both resting state fMRI (R-fMRI) and task-based fMRI (T-fMRI) have been widely used to study the functional activities of the human brain during task-free and task-performance periods, respectively. However, due to the difficulty in strictly controlling the participating subject’s mental status and their cognitive behaviors during fMRI scans, it has been very challenging to tell whether or not an R-fMRI/T-fMRI scan truly reflects the participant’s functional brain states in task-free/task-performance. This paper presents a novel approach to characterizing the brain’s functional status into task-free or task-performance states. The basic idea here is that the brain’s functional state is represented by a whole-brain quasi-stable connectivity pattern (WQCP), and an effective sparse coding procedure was then applied to learn the atomic connectivity patterns (ACP) of both task-free and task-performance states based on training R-fMRI and T-fMRI data. Our experimental results demonstrated that the learned ACPs for R-fMRI and T-fMRI datasets are substantially different, as expected. However, a certain portion of ACPs from R-fMRI and T-fMRI datasets are overlapping, suggesting that those subjects with overlapping ACPs were not in the expected task-free/task-performance states during R-fMRI/T-fMRI scans.

Modern statistical inference techniques may be able to improve the sensitivity and specificity of resting state functional MRI (rs-fMRI) connectivity analysis through more realistic characterization of distributional assumptions. In simulation, the advantages of such modern methods are readily demonstrable. However quantitative

empirical

validation remains elusive

in vivo

as the true connectivity patterns are unknown and noise/artifact distributions are challenging to characterize with high fidelity. Recent innovations in capturing finite sample behavior of asymptotically consistent estimators (i.e., SIMulation and EXtrapolation - SIMEX) have enabled direct estimation of bias given single datasets. Herein, we leverage the theoretical core of SIMEX to study the properties of inference methods in the face of diminishing data (in contrast to increasing noise). The stability of inference methods with respect to

synthetic

loss of

empirical

data (defined as

resilience

) is used to quantify the empirical performance of one inference method relative to another. We illustrate this new approach in a comparison of ordinary and robust inference methods with rs-fMRI.

Pathologies like autism and schizophrenia are a broad set of disorders with multiple etiologies in the same diagnostic category. This paper presents a method for unsupervised cluster analysis using multi-edge similarity graphs that combine information from different modalities. The method alleviates the issues with traditional supervised classification methods that use diagnostic labels and are therefore unable to exploit or elucidate the underlying heterogeneity of the dataset under analysis. The framework introduced in this paper has the ability to employ diverse features that define different aspects of pathology obtained from different modalities to create a multi-edged graph on which clustering is performed. The weights on the multiple edges are optimized using a novel concept of ‘holding power’ that describes the certainty with which a subject belongs to a cluster. We apply the technique to two separate clinical populations of autism spectrum disorder (ASD) and schizophrenia (SCZ), where the multi-edged graph for each population is created by combining information from structural networks and cognitive scores. For the ASD-control population the method clusters the data into two classes and the SCZ-control population is clustered into four. The two classes in ASD agree with underlying diagnostic labels with 92% accuracy and the SCZ clustering agrees with 78% accuracy, indicating a greater heterogeneity in the SCZ population.

Homogenising the availability of manually generated information in large databases has been a key challenge of medical imaging for many years. Due to the time consuming nature of manually segmenting, parcellating and localising landmarks in medical images, these sources of information tend to be scarce and limited to small, and sometimes morphologically similar, subsets of data. In this work we explore a new framework where these sources of information can be propagated to morphologically dissimilar images by diffusing and mapping the information through intermediate steps. The spatially variant data embedding uses the local morphology and intensity similarity between images to diffuse the information only between locally similar images. This framework can thus be used to propagate any information from any group of subject to every other subject in a database with great accuracy. Comparison to state-of-the-art propagation methods showed highly statistically significant (

p

 < 10

− 4

) improvements in accuracy when propagating both structural parcelations and brain segmentations geodesically.

In-vivo parcellation of the cerebral cortex via non-invasive neuroimaging data has been in active research for years. A variety of model-driven and/or data-driven computational approaches have been proposed to parcellate the cortex. However, two fundamental common issues in these parcellation methodologies are the features or attributes used to define boundaries between cortical regions and the establishment of correspondences of the parcellated regions across different brains. This paper uses a novel DTI-derived fiber shape feature for the parcellation of cortical gyrus into fine-granularity segments. The gyral parcellation is formulated and solved as a surface vertex clustering problem, in which fiber shape feature similarity is used to define the distances between vertices. Then, we designed and applied a novel multi-view spectral clustering algorithm to group the vertices into group-wise consistent gyral segments across different brains. The experimental results showed that the precentral and postcentral gyrus, as two test-beds, can be consistently parcellated into 10 segments on both hemispheres across different subjects. Evaluation studies using benchmark task-based fMRI and cortical landmarks demonstrated the effectiveness and validity of the proposed methods.

The ensemble average propagator (EAP) describes the 3D average diffusion process of water molecules, capturing both its radial and angular contents, and hence providing rich information about complex tissue microstructure properties. Bessel Fourier orientation reconstruction (BFOR) is one of several analytical, non-Cartesian EAP reconstruction schemes employing multiple shell acquisitions that have recently been proposed. Such modeling bases have not yet been fully exploited in the extraction of rotationally invariant

q

-space indices that describe the degree of diffusion anisotropy/restrictivity. Such quantitative measures include the zero-displacement probability (

Po

), mean squared displacement (MSD),

q

-space inverse variance (QIV), and generalized fractional anisotropy (GFA), and all are simply scalar features of the EAP. In this study, a general relationship between MSD and

q

-space diffusion signal is derived and an EAP-based definition of GFA is introduced. A significant part of the paper is dedicated to utilizing BFOR in a clinical dataset, comprised of 5 multiple sclerosis (MS) patients and 4 healthy controls, to estimate

Po

, MSD, QIV, and GFA of corpus callosum, and specifically, to see if such indices can detect changes between normal appearing white matter (NAWM) and healthy white matter (WM). Although the sample size is small, this study is a proof of concept that can be extended to larger sample sizes in the future.

Diffusion spectrum imaging (DSI) from multiple diffusion-weighted images (DWI) allows to image the complex geometry of water diffusion in biological tissue. To capture the structure of DSI data, we propose to use sparse coding constrained by physical properties of the signal, namely symmetry and positivity, to learn a dictionary of diffusion profiles. Given this estimated model of the signal, we can extract better estimates of the signal from noisy measurements and also speed up acquisition by reducing the number of acquired DWI while giving access to high resolution DSI data. The method learns jointly for all the acquired DWI and scales to full brain data. Working with two sets of 515 DWI images acquired on two different subjects we show that using just half of the data (258 DWI) we can better predict the other 257 DWI than the classic symmetry procedure. The observation holds even if the diffusion profiles are estimated on a different subject dataset from an undersampled q-space of 40 measurements.

Diffusion-sensitized magnetic resonance imaging provides information about the fibrous structure of the human brain. However, this information is not sufficient to reconstruct the underlying fiber network, because the nature of diffusion provides only conditional fiber densities. That is, it is possible to infer the percentage of bundles that pass a voxel with a certain direction, but the absolute number of fibers is inaccessible. In this work we propose a conservation equation for tensor fields that can infer this number up to a factor. Simulations on synthetic phantoms show that the approach is able to derive the densities correctly for various configurations. In-vivo results on 20 healthy volunteers are plausible and consistent, while a rigorous evaluation is difficult, because conclusive data from both MRI and histology remain elusive even on the most studied brain structures.

Diffusion MRI measures micron scale displacement of water molecules, providing unique insight into microstructural tissue architecture. However, current practical image resolution is in the millimeter scale, and thus diffusivities from many tissue compartments are averaged in each voxel, reducing the sensitivity and specificity of the measurement to subtle pathologies. Recent studies have pointed out that eliminating the contribution of extracellular water increases the sensitivity of the diffusion measures to tissue architecture. Moreover, in brain imaging, estimation of the extracellular volume appears to indicate pathological processes such as atrophy, edema and neuroinflammation. Here we study the free-water method, which assumes a bi-tensor model. We add low b-value shells to a regular DTI acquisition and present methods to improve the estimation of the model parameters using the extra information. In addition, we define a Laplace-Beltrami regularization operator that further stabilizes the multi-shell estimation.

In High Angular Resolution Diffusion Imaging (HARDI), Orientation Distribution Function (ODF) and Ensemble Average Propagator (EAP) are two important Probability Density Functions (PDFs) which reflect the water diffusion and fiber orientations. Spherical Polar Fourier Imaging (SPFI) is a recent model-free multi-shell HARDI method which estimates both EAP and ODF from the diffusion signals with multiple

b

values. As physical PDFs, ODFs and EAPs are nonnegative definite respectively in their domains

$\mathbb{S}^2$

and ℝ

3

. However, existing ODF/EAP estimation methods like SPFI seldom consider this natural constraint. Although some works considered the nonnegative constraint on the given discrete samples of ODF/EAP, the estimated ODF/EAP is not guaranteed to be nonnegative definite in the whole continuous domain. The Riemannian framework for ODFs and EAPs has been proposed via the square root parameterization based on pre-estimated ODFs and EAPs by other methods like SPFI. However, there is no work on how to estimate the square root of ODF/EAP called as the wavefuntion directly from diffusion signals. In this paper, based on the Riemannian framework for ODFs/EAPs and Spherical Polar Fourier (SPF) basis representation, we propose a unified model-free multi-shell HARDI method, named as Square Root Parameterized Estimation (SRPE), to simultaneously estimate both the wavefunction of EAPs and the nonnegative definite ODFs and EAPs from diffusion signals. The experiments on synthetic data and real data showed SRPE is more robust to noise and has better EAP reconstruction than SPFI, especially for EAP profiles at large radius.

Current methods in high angular resolution diffusion imaging (HARDI) estimate the probability density function of water diffusion as a continuous-valued orientation distribution function (ODF) on the sphere. However, such methods could produce an ODF with negative values, because they enforce non-negativity only at finitely many directions. In this paper, we propose to enforce non-negativity on the continuous domain by enforcing the positive semi-definiteness of Toeplitz-like matrices constructed from the spherical harmonic representation of the ODF. We study the distribution of the eigenvalues of these matrices and use it to derive an iterative semi-definite program that enforces non-negativity on the continuous domain. We illustrate the performance of our method and compare it to the state-of-the-art with experiments on synthetic and real data.

Registration of DWI data, unlike scalar image data, is complicated by the need of reorientation algorithms for keeping the orientation architecture of each voxel aligned with the rest of the image. This paper presents an algorithm for effective and efficient warping and reconstruction of diffusion-weighted imaging (DWI) signals for the purpose of spatial transformation. The key idea is to decompose the DWI signal profile, a function defined on a unit sphere, into a series of weighted fiber basis functions (FBFs), reorient these FBFs independently based on the local affine transformation, and then recompose the reoriented FBFs to obtain the final transformed DWI signal profile. We enforce a sparsity constraint on the weights of the FBFs during the decomposition to reflect the fact that the DWI signal profile typically gains its information from a limited number of fiber populations. A non-negative constraint is further imposed so that noise-induced negative lobes in the profile can be avoided. The proposed framework also explicitly models the isotropic component of the diffusion signals to avoid undesirable reorientation artifacts in signal reconstruction. In contrast to existing methods, the current algorithm is executed directly in the DWI signal space, thus allowing any diffusion models to be fitted to the data after transformation.

It’s well known that in diffusion MRI (dMRI), fibre crossing is an important problem for most existing diffusion tensor imaging (DTI) based tractography algorithms. To overcome these limitations, High Angular Resolution Diffusion Imaging (HARDI) based tractography has been proposed with a particular emphasis on the the Orientation Distribution Function (ODF). In this paper, we advocate the use of the Ensemble Average Propagator (EAP) instead of the ODF for tractography in dMRI and propose an original and efficient EAP-based tractography algorithm that outperforms the classical ODF-based tractography, in particular, in the regions that contain complex fibre crossing configurations. Various experimental results including synthetic, phantom and real data illustrate the potential of the approach and clearly show that our method is especially efficient to handle regions where fiber bundles are crossing, and still well handle other fiber bundle configurations such as U-shape and kissing fibers.

Segmentation of lungs with large tumors is a challenging and time-consuming task, especially for 4D CT data sets used in radiation therapy. Existing lung segmentation methods are ineffective in these cases, because they are either not able to deal with large tumors and/or process every 3D image independently neglecting temporal information. In this paper, we present a approach for model-based 4D segmentation of lungs with large tumors in 4D CT data sets. In our approach, a 4D statistical shape model that accounts for inter- and intra-patient variability is fitted to the 4D image sequence, and the segmentation result is refined by a 4D graph-based optimal surface finding. The approach is evaluated using 10 4D CT data sets of lung tumor patients. The segmentation results are compared with a standard intensity-based approach and a 3D version of the presented model-based segmentation method. The intensity-based approach shows a better performance for normal lungs, however, fails in presence of large lung tumors. Although overall performance of 3D and 4D model-based segmentation is similar, the results indicate improved temporal coherence and improved robustness with respect to the segmentation parameters for the 4D model-based segmentation.

While Markov random fields are very popular segmentation models in medical image processing, the associated maximum a posteriori (MAP) estimation problem is usually solved using iterative methods that are prone to local maxima. We show that a variant of the random walker algorithm can be seen as a relaxation method for the MAP problem under the Potts model. The key advantage of this technique is that it boils down to a sparse linear system with a uniquely defined explicit solution. Our experiments further demonstrate that the resulting MAP approximation can be used to improve the classical mean-field algorithm in terms of MAP estimation quality.

Fuzzy boundaries of anatomical structures in medical images make segmentation a challenging task. We present a new segmentation method that addresses the fuzzy boundaries problem. Our method maps the lengths of 3D rays cast from a seed point to the unit sphere, estimates the fuzzy boundaries location by thresholding the gradient magnitude of the rays lengths, and derives the true boundaries by Laplacian interpolation on the sphere. Its advantages are that it does not require a global shape prior or curvature based constraints, that it has an automatic stopping criteria, and that it is robust to anatomical variability, noise, and parameters values settings. Our experimental evaluation on 23 segmentations of kidneys and on 16 segmentations of abdominal aortic aneurysms (AAA) from CT scans yielded an average volume overlap error of 12.6% with respect to the ground-truth. These results are comparable to those of other segmentation methods without their underlying assumptions.

Pectoral muscle segmentation is an important step in automatic breast image analysis methods and crucial for multi-modal image registration. In breast MRI, accurate delineation of the pectoral is important for volumetric breast density estimation and for pharmacokinetic analysis of dynamic contrast enhancement. In this paper we propose and study the performance of atlas-based segmentation methods evaluating two fully automatic breast MRI dedicated strategies on a set of 27 manually segmented MR volumes. One uses a probabilistic model and the other is a multi-atlas registration based approach. The multi-atlas approach performed slightly better, with an average Dice coefficient (DSC) of 0.74, while with the much faster probabilistic method a DSC of 0.72 was obtained.

The detection of gad-enhancing lesions in brain MRI of Multiple Sclerosis (MS) patients is of great interest since they are important markers of disease activity. However, many of the enhancing voxels are associated with normal structures (i.e. blood vessels) or noise in the MRI, making the detection of gad-enhancing lesions a challenging task. Furthermore, these lesions are typically small and in close proximity to vessels. In this paper, we present an automatic, probabilistic Hierarchical Conditional Random Field (HCRF) framework for detection of gad-enhancing lesions in brain images of patients with MS. In the first level, a CRF with unary and pairwise potentials is used to identify candidate lesion voxel. In the second level, these lesion candidates are grouped based on anatomical and spatial features, and feature-specific lesion based CRF models are designed for each group. This lesion level CRF incorporates higher order potentials which account for shape, group intensities and symmetries. The proposed algorithm is trained on 92 multimodal clinical datasets acquired from Relapsing-Remitting MS patients during multicenter clinical trials and is evaluated on 30 independent cases. The experimental results show a sensitivity of 98%, a positive predictive value of 66% and an average false positive count of 1.55, outperforming the CRF and MRF frameworks proposed in [1].

Image segmentation plays a crucial role in many medical imaging applications by automatically locating the regions of interest. Typically supervised learning based segmentation methods require a large set of accurately labeled training data. However, thel labeling process is tedious, time consuming and sometimes not necessary. We propose a robust logistic regression algorithm to handle label outliers such that doctors do not need to waste time on precisely labeling images for training set. To validate its effectiveness and efficiency, we conduct carefully designed experiments on cervigram image segmentation while there exist label outliers. Experimental results show that the proposed robust logistic regression algorithms achieve superior performance compared to previous methods, which validates the benefits of the proposed algorithms.

In this paper, we present a liver segmentation approach. In which, the relation between neighboring slices in CT images is utilized to estimate shape and statistical information of the liver. This information is then integrated with the graph cuts algorithm to segment the liver in each CT slice. This approach does not require prior models construction, and it uses single phase CT images; even so, it is talented to deal with complex shape and intensity variations. Moreover, it eliminates the burdens associated with model construction like data collection, manual segmentation, registration, and landmark correspondence. In contrast, it requires a low user interaction to determine the liver landmarks on a single CT slice only. The proposed approach has been evaluated on 10 CT images with several liver abnormalities, including tumors and cysts, and it achieved high average scores of 81.7 using MICCAI-2007 Grand Challenge scoring system. Compared to contemporary approaches, our approach requires significantly less interaction and processing time.

An emerging topic is to build image segmentation systems that can segment hundreds to thousands of objects (i.e. cell segmentation \ tracking, full brain parcellation, full body segmentation, etc.). Multi-object Level Set Methods (MLSM) perform this task with the benefit of sub-pixel precision. However, current implementations of MLSM are not as computationally or memory efficient as their region growing and graph cut counterparts which lack sub-pixel precision. To address this performance gap, we present a novel parallel implementation of MLSM that leverages the sparse properties of the algorithm to minimize its memory footprint for multiple objects. The new method, Multi-Object Geodesic Active Contours (MOGAC), can represent N objects with just two functions: a label mask image and unsigned distance field. The time complexity of the algorithm is shown to be O((M^d)/P) for M^d pixels and P processing units in dimension d={2,3}, independent of the number of objects. Results are presented for 2D and 3D image segmentation problems.

Zonal segmentation of the prostate into the central gland and peripheral zone is a useful tool in computer-aided detection of prostate cancer, because occurrence and characteristics of cancer in both zones differ substantially. In this paper we present a pattern recognition approach to segment the prostate zones. It incorporates three types of features that can differentiate between the two zones: anatomical, intensity and texture. It is evaluated against a multi-parametric multi-atlas based method using 48 multi-parametric MRI studies. Three observers are used to assess inter-observer variability and we compare our results against the state of the art from literature. Results show a mean Dice coefficient of 0.89 ± 0.03 for the central gland and 0.75 ± 0.07 for the peripheral zone, compared to 0.87 ± 0.04 and 0.76 ± 0.06 in literature. Summarizing, a pattern recognition approach incorporating anatomy, intensity and texture has been shown to give good results in zonal segmentation of the prostate.

Cochlear implant (CI) surgery is considered standard of care treatment for severe hearing loss. CIs are currently programmed using a one-size-fits-all type approach. Individualized, position-based CI programming schemes have the potential to significantly improve hearing outcomes. This has not been possible because the position of stimulation targets is unknown due to their small size and lack of contrast in CT. In this work, we present a segmentation approach that relies on a weighted active shape model created using microCT scans of the cochlea acquired ex-vivo in which stimulation targets are visible. The model is fitted to the partial information available in the conventional CTs and used to estimate the position of structures not visible in these images. Quantitative evaluation of our method results in Dice scores averaging 0.77 and average surface errors of 0.15 mm. These results suggest that our approach can be used for position-dependent image-guided CI programming methods.

Most image segmentation algorithms are designed to estimate a single segmentation for each image, where the gold standard segmentation is often labeled by a human expert. However, it is common that multiple manual segmentations are available for some images, e.g. independently labeled by different experts. For efficient usages of manual segmentations, we propose to simultaneously produce automatic estimations for each expert. The key advantage of this proposal is that it allows to incorporate the correlations between different experts to improve the accuracy of automatic segmentation. In a brain image segmentation problem, where for each image six manual segmentations are available, we show that jointly estimating several manual segmentations produces significant improvement over independently estimating each of them.

In this paper we propose an atlas-aided probabilistic model-based segmentation method for estimating the fibroglandular tissue in breast MRI, where a novel fibroglandular tissue atlas is learned to aid the segmentation. The atlas represents a pixel-wise likelihood of being fibroglandular tissue in the breast, which is derived by combining deformable image warping, using aligned breast contour points as landmarks, with a kernel density estimation technique. A mixture multivariate model is learned to characterize the breast tissue using MR image features, and the segmentation is subsequently based on examining the posterior probability where the learned atlas is incorporated as the prior probability. In our experiments, the algorithm-generated segmentation results of 10 cases are compared to the manual segmentations, verified by an experienced breast imaging radiologist, to assess the accuracy of the algorithm, where the Dice’s Similarity Coefficient (DSC) shows a 0.85 agreement. The proposed automated segmentation method could be used to estimate the volumetric amount of fibroglandular tissue in the breast for breast cancer risk estimation.

Severe cases of spinal deformities such as scoliosis are usually treated by a surgery where instrumentation (hooks, screws and rods) is installed to the spine to correct deformities. Even if the purpose is to obtain a normal spine curve, the result is often straighter than normal. In this paper, we propose a fast statistical reconstruction algorithm based on a general model which can deal with such instrumented spines. To this end, we present the concept of multilevel statistical model where the data are decomposed into a within-group and a between-group component. The reconstruction procedure is formulated as a second-order cone program which can be solved very fast (few tenths of a second). Reconstruction errors were evaluated on real patient data and results showed that multilevel modeling allows better 3D reconstruction than classical models.

Intravascular ultrasound (IVUS) is a catheter-based medical imaging technique that produces cross-sectional images of blood vessels. In this paper, we present a method for the segmentation of the luminal border using IVUS radio frequency (RF) data. Specifically, we parameterize the lumen contour using Fourier series. This contour is deformed by minimizing a cost function that is formulated using a probabilistic approach in which the

a priori

term is obtained using the prediction confidence of a Support Vector Machine classifier using features extracted from the RF signal. We evaluated the performance of our method by comparing our results with manual segmentations from two expert observers on 280 frames from eight 40 MHz IVUS sequences from rabbits and pigs. The performance was evaluated using the Dice similarity coefficient, coefficient of determination, and linear regressions of the lumen area for each frame. Our results indicate the feasibility of our method for the segmentation of the lumen from IVUS RF data.

In this paper, we present a novel method by incorporating information theory into the learning-based approach for automatic and accurate pelvic organ segmentation (including the prostate, bladder and rectum). We target 3D CT volumes that are generated using different scanning protocols (e.g., contrast and non-contrast, with and without implant in the prostate, various resolution and position), and the volumes come from largely diverse sources (e.g., diseased in different organs). Three key ingredients are combined to solve this challenging segmentation problem. First, marginal space learning (MSL) is applied to efficiently and effectively localize the multiple organs in the largely diverse CT volumes. Second, learning techniques, steerable features, are applied for robust boundary detection. This enables handling of highly heterogeneous texture pattern. Third, a novel information theoretic scheme is incorporated into the boundary inference process. The incorporation of the Jensen-Shannon divergence further drives the mesh to the best fit of the image, thus improves the segmentation performance. The proposed approach is tested on a challenging dataset containing 188 volumes from diverse sources. Our approach not only produces excellent segmentation accuracy, but also runs about eighty times faster than previous state-of-the-art solutions. The proposed method can be applied to CT images to provide visual guidance to physicians during the computer-aided diagnosis, treatment planning and image-guided radiotherapy to treat cancers in pelvic region.

In this article, the authors demonstrate that you can improve the performance of the registration of a point distribution model (PDM) by accurately estimating the structure of an undirected graphical model that represents the statistical shape model (SSM) of a target surface. Many existing methods for constructing SSMs determine the structure of the graphical model without analyzing the conditional dependencies among the points in PDM, though an edge in the PDM should link two nodes if and only if they are conditionally dependent. In this study, the authors employed four popular methods for estimating the structure of graphical model and obtained four different SSMs from an identical set of training surfaces. The registration performances of the SSMs were experimentally compared, and the results showed that the graphical lasso, which could estimate more accurate structure of the graphical model by avoiding the overfitting to the training data, outperformed the other methods.

This paper presents a method for automatically estimating the quality of Parasternal Long AXis (PLAX) B-mode echocardiograms. The purpose of the algorithm is to provide live feedback to the user on the quality of the acquired image. The proposed approach uses Generalized Hough Transform to compare the structures derived from the incoming image to a representative atlas, thereby providing a quality metric (PQM). On 133 PLAX images from 35 patients, we show: 1) PQM has high correlation with manual ratings from an expert echocardiographer 2) PQM has high correlation with contrast-to-noise ratio, a traditional indicator of image quality 3) on images with high PQM, error in automatic septal wall thickness measurement is low, and

vice versa

.

Coarctation of the aorta (CoA), is a congenital defect characterized by a severe narrowing of the aorta, usually distal to the aortic arch. The treatment options include surgical repair, stent implantation, and balloon angioplasty. In order to evaluate the physiological significance of the pre-operative coarctation and to assess the post-operative results, the hemodynamic analysis is usually performed by measuring the pressure gradient (

$\triangle P$

) across the coarctation site via invasive cardiac catheterization. The measure of success is reduction of the (

$\triangle P > 20 mmHg$

) systolic blood pressure gradient. In this paper, we propose a non-invasive method based on Computational Fluid Dynamics and MR imaging to estimate the pre- and post-operative hemodynamics for both native and recurrent coarctation patients. High correlation of our results and catheter measurements is shown on corresponding pre- and post-operative examination of 5 CoA patients.

Purpose:

Various methods exist for interpolating diffusion tensor fields, but none of them linearly interpolate tensor shape attributes. Linear interpolation is expected not to introduce spurious changes in tensor shape.

Methods:

Herein we define a new linear invariant (LI) tensor interpolation method that linearly interpolates components of tensor shape (tensor invariants) and recapitulates the interpolated tensor from the linearly interpolated tensor invariants and the eigenvectors of a linearly interpolated tensor. The LI tensor interpolation method is compared to the Euclidean (EU), affine-invariant Riemannian (AI), log-Euclidean (LE) and geodesic-loxodrome (GL) interpolation methods using both a synthetic tensor field and three experimentally measured cardiac DT-MRI datasets.

Results:

EU, AI, and LE introduce significant microstructural bias, which can be avoided through the use of GL or LI.

Conclusion:

GL introduces the least microstructural bias, but LI tensor interpolation performs very similarly and at substantially reduced computational cost.

The complex morphological structure of the left ventricular endocardial surface and its relation to the severity of arterial stenosis has not yet been thoroughly investigated due to the limitations of conventional imaging techniques. By exploiting the recent developments in Multirow-Detector Computed Tomography (MDCT) scanner technology, the complex endocardial surface morphology of the left ventricle is studied and the cardiac segments affected by coronary arterial stenosis localized via analysis of Computed Tomography (CT) image data obtained from a 320-MDCT scanner. The non-rigid endocardial surface data is analyzed using an isometry-invariant

Bag-of-Words

(BOW) feature-based approach. The clinical significance of the analysis in identifying, localizing and quantifying the incidence and extent of coronary artery disease is investigated. Specifically, the association between the incidence and extent of coronary artery disease and the alterations in the endocardial surface morphology is studied. The results of the proposed approach on 15 normal data sets, and 12 abnormal data sets exhibiting coronary artery disease with varying levels of severity are presented. Based on the characterization of the endocardial surface morphology using the

Bag-of-Words

features, a neural network-based classifier is implemented to test the effectiveness of the proposed morphological analysis approach. Experiments performed on a strict leave-one-out basis are shown to exhibit a distinct pattern in terms of classification accuracy within the cardiac segments where the incidence of coronary arterial stenosis is localized.

In current clinical practice, examinations of the carotid artery bifurcation are commonly carried out with Computed Tomography Angiography,(CTA) or contrast-enhanced Magnetic Resonance Angiography (ce,MRA). Quantitative information about vessel morphology, extracted from segmentations, is promising for diagnosis of vessel pathologies. However, both above-mentioned techniques require the administration of contrast media. In contrary, non-ce MRA methods such as Time-of-Flight,(TOF) provide fully non-invasive imaging without any exogenous contrast agent. The diagnostic value of TOF MRA, however, for assessment of the carotid bifurcation area can be hampered due to its susceptibility to irregular blood flow patterns. Conventional methods for lumen segmentation are very sensitive to such signal voids and produce inaccurate results. In this work, a novel, fully automatic 3D segmentation algorithm is proposed which uses prior knowledge about irregular flow patterns. The presented technique has been successfully tested on eleven volunteer datasets as well as in a patient case, offering the comparison to CTA images. The sensitivity could be increased by 29.2% to 85.6% compared to standard level set methods. The root mean squared error in diameter measurements was reduced from 4.85,mm to 1.44,mm.

While Magnetic Resonance Imaging is capable of separating water and fat components in the body, mapping of magnetic field inhomogeneities is essential for the successful application of this process. In this study, we address the problem of field map estimation using a convex-relaxed max-flow method. We propose a novel two-stage approach that leads to the global optimum of the proposed problem. The first stage minimizes the signal residuals via a convex-relaxed minimum description length (MDL)-based approach. The MDL-based labeling model penalizes the total number of appearing labels, which helps to avoid field map errors when abrupt changes in field homogeneity exist. By exploring the whole range of possible frequency offsets, this stage ensures limiting the estimated field offset within certain boundaries where the global minimum resides. The second stage employs the output of the labeling model in a commonly used gradient-descent based method (known as IDEAL) to converge to the exact global minimum, i.e. the required value of the field offset. Experimental results for cardiac imaging, where challenging field inhomogeneities exist, showed that our method significantly outperforms over a widely-used technique for fat/water separation in terms of robustness and efficiency.

This study investigates regional heart motion abnormality detection via multiview fusion in cine cardiac MR images. In contrast to previous methods which rely only on short-axis image sequences, the proposed approach exploits the information from several other long-axis image sequences, namely, 2-chamber, 3-chamber and 4-chamber MR images. Our analysis follows the standard issued by American Heart Association to identify 17 standardized left ventricular segments. The proposed method first computes an initial sequence of corresponding myocardial points using a nonrigid image registration algorithm within each sequence. Then, these points were mapped to 3D space and tracked using Unscented Kalman Filter (UKS). We propose a maximum likelihood based track-to-track fusion approach to combine UKS tracks from multiple image views. Finally, we use a Shannon’s differential entropy of distributions of potential classifiers obtained from multiview fusion estimates, and a naive Bayes classifier algorithm to automatically detect abnormal functional regions of the myocardium. We proved the benefits of the proposed method by comparing the classification results with and without fusion over 480 regional myocardial segments obtained from 30 subjects. The evaluations in comparisons to the ground truth classifications by radiologists showed that the proposed fusion yielded an area-under-the-curve (AUC) of 95.9%, bringing a significant improvement of 3.8 % in comparisons to previous methods that use only short-axis images.

The cardiac ejection fraction (EF) depends on the volume variation of the left ventricle (LV) cavity during a cardiac cycle, and is an essential measure in the diagnosis of cardiovascular diseases. It is often estimated via manual segmentation of several images in a cardiac sequence, which is prohibitively time consuming, or via automatic segmentation, which is a challenging and computationally expensive task that may result in high estimation errors. In this study, we propose to estimate the EF in real-time directly from image statistics using machine learning technique. From a simple user input in only one image, we build for all the images in a subject dataset (200 images) a statistic based on the Bhattacharyya coefficient of similarity between image distributions. We demonstrate that these statistics are non-linearly related to the LV cavity areas and, therefore, can be used to estimate the EF via an Artificial Neural Network (ANN) directly. A comprehensive evaluation over 20 subjects demonstrated that the estimated EFs correlate very well with those obtained from independent manual segmentations.

New minimal-invasive interventions such as transcatheter valve procedures exploit multiple imaging modalities to guide tools (fluoroscopy) and visualize soft tissue (transesophageal echocardiography (TEE)). Currently, these complementary modalities are visualized in separate coordinate systems and on separate monitors creating a challenging clinical workflow. This paper proposes a novel framework for fusing TEE and fluoroscopy by detecting the pose of the TEE probe in the fluoroscopic image. Probe pose detection is challenging in fluoroscopy and conventional computer vision techniques are not well suited. Current research requires manual initialization or the addition of fiducials. The main contribution of this paper is autonomous six DoF pose detection by combining discriminative learning techniques with a fast binary template library. The pose estimation problem is reformulated to incrementally detect pose parameters by exploiting natural invariances in the image. The theoretical contribution of this paper is validated on synthetic, phantom and

in vivo

data. The practical application of this technique is supported by accurate results (< 5 mm in-plane error) and computation time of 0.5s.

Interventional guidance systems require surgical navigation systems to register different tools and devices together. Standard navigation systems have various drawbacks leading to target registration errors (TRE) of around 3mm. The aim of this work is to introduce the photoacoustic (PA) effect as a direct 3D ultrasound (US) to video registration method. We present our experimental setup and demonstrate its feasibility on both a synthetic phantom and an

ex vivo

tissue phantom. We achieve an average TRE of 560 microns and standard deviation of 280 microns on a synthetic phantom. Also, an average TRE of 420 microns and standard deviation of 150 microns on the

ex vivo

tissue phantom are obtained. We describe a roadmap to bring this system into the surgical setting and highlight possible challenges along the way.

Despite increased use of robotic catheter navigation systems for endovascular intervention procedures, current master-slave platforms have not yet taken into account dexterous manipulation skill used in traditional catheterization procedures. Information on tool forces applied by operators is often limited. A novel force/torque sensor is developed in this paper to obtain behavioural data across different experience levels and identify underlying factors that affect overall operator performance. The miniature device can be attached to any part of the proximal end of the catheter, together with a position sensor attached to the catheter tip, for relating tool forces to catheter dynamics and overall performance. The results show clear differences in manipulation skills between experience groups, thus providing insights into different patterns and range of forces applied during routine endovascular procedures. They also provide important design specifications for ergonomically optimized catheter manipulation platforms with added haptic feedback while maintaining natural skills of the operators.

In the context of retinal microsurgery, visual tracking of instruments is a key component of robotics assistance. The difficulty of the task and major reason why most existing strategies fail on

in-vivo

image sequences lies in the fact that complex and severe changes in instrument appearance are challenging to model. This paper introduces a novel approach, that is both data-driven and complementary to existing tracking techniques. In particular, we show how to learn and integrate an accurate detector with a simple gradient-based tracker within a robust pipeline which runs at framerate. In addition, we present a fully annotated dataset of retinal instruments in

in-vivo

surgeries, which we use to quantitatively validate our approach. We also demonstrate an application of our method in a laparascopy image sequence.

Over the last years, range imaging (RI) techniques have been proposed for patient positioning and respiration analysis in motion compensation. Yet, current RI based approaches for patient positioning employ rigid-body transformations, thus neglecting free-form deformations induced by respiratory motion. Furthermore, RI based respiration analysis relies on non-rigid registration techniques with run-times of several seconds. In this paper we propose a real-time framework based on RI to perform respiratory motion compensated positioning and non-rigid surface deformation estimation in a joint manner. The core of our method are pre-procedurally obtained 4-D shape priors that drive the intra-procedural alignment of the patient to the reference state, simultaneously yielding a rigid-body table transformation and a free-form deformation accounting for respiratory motion. We show that our method outperforms conventional alignment strategies by a factor of 3.0 and 2.3 in the rotation and translation accuracy, respectively. Using a GPU based implementation, we achieve run-times of 40 ms.

We propose novel methods for (a) detection of a catheter in fluoroscopic images and (b) reconstruction of this catheter from two views. The novelty of (a) is a reduced user interaction and a higher accuracy. It requires only a single seed point on the catheter in the fluoroscopic image. Using this starting point, possible parts of the catheter are detected using a graph search. An evaluation of the detection using 66 clinical fluoroscopic images yielded an average error of 0.7 mm ± 2.0 mm. The novelty of (b) is a better ability to deal with highly curved objects as it selects an optimal set of point correspondences from two point sequences describing the catheters in two fluoroscopic images. The selected correspondences are then used for computation of the 3-D reconstruction. The evaluation on 33 clinical biplane images yielded an average backprojection error of 0.4 mm ± 0.6 mm.

Tool tracking is an accepted capability for computer-aided surgical intervention which has numerous applications, both in robotic and manual minimally-invasive procedures. In this paper, we describe a tracking system which learns visual feature descriptors as class-specific landmarks on an articulated tool. The features are localized in 3D using stereo vision and are fused with the robot kinematics to track all of the joints of the dexterous manipulator. Experiments are performed using previously-collected porcine data from a surgical robot.

We present image-based methods for tracking teeth in a video image with respect to a CT scan of the jaw, in order to enable a novel light-weight augmented reality (AR) system in orthodontistry. Its purpose is guided bracket placement in orthodontic correction. In this context, our goal is to determine the position of the patient maxilla and mandible in a video image solely based on a CT scan. This is suitable for image guidance through an overlay of the video image with the planned position of brackets in a monocular AR system. Our tracking algorithm addresses the contradicting requirements of robustness, accuracy and performance in two problem-specific formulations. First, we exploit a distance-based modulation of two iso-surfaces from the CT image to approximate the appearance of the gum line. Second, back-projection of previous video frames to an iso-surface is used to account for recently placed brackets. In combination, this novel algorithm allowed us to track several sequences of three patient videos of real procedures, despite difficult lighting conditions. Paired with a systematic evaluation, we were able to show practical feasibility of such a system.

The alignment of the lower limb in high tibial osteotomy (HTO) or total knee arthroplasty (TKA) must be determined intraoperatively. One way to do so is to deform the mechanical axis deviation (MAD), for which a tolerance measurement of 10mm is widely accepted. Many techniques are proposed in clinical practice such as visual inspection, cable method, grid with lead impregnated reference lines, or more recently, navigation systems. Each has their disadvantages including reliability of the MAD measurement, excess radiation, prolonged operation time, complicated setup and high cost. To alleviate such shortcomings, we propose a novel clinical protocol that allows quick and accurate intraoperative calculation of MAD. This is achieved by an X-ray stitching method requiring only three X-ray images placed into a panoramic image frame during the entire procedure. The method has been systematically analyzed in a simulation framework in order to investigate its accuracy and robustness. Furthermore, we validated our protocol via a preclinical study comprising 19 human cadaver legs. Four surgeons determined MAD measurements using our X-ray panorama and compared these values to a gold-standard CT-based technique. The maximum average MAD error was 3.5mm which shows great potential for the technique.

In this article an ultrasound elastography technology is reported which provides quantitative images of tissue elasticity from deep soft tissue. The technique is analogous to Magnetic Resonance Elastography in the use of external mechanical vibrations which can penetrate deep tissue. Multifrequency steady-state mechanical vibrations are applied to the tissue at the skin and tissue displacements are measured by a conventional ultrasound system. Absolute values of tissue elasticity are computed in real-time for each frequency and displayed to the physician. The quantitative elasticity images produced by the technology are validated with magnetic resonance elastography images as the gold standard on standard elasticity phantoms. Preliminary

in-vivo

data from healthy volunteers are presented which show the potential of the technology for clinical use. The system is currently being used in different clinical studies to image kidney fibrosis, liver fibrosis, and prostate cancer.

The ability to find image similarities (feature matching) between laparoscopic views is essential in many robotic-assisted Minimally-Invasive Surgery (MIS) applications. Differently from feature tracking methods, feature matching does not make any restrictive assumption about the sequential nature of the two images or about the organ motion, and could then be used, e.g., to recover tracked features that were lost due to a prolonged occlusion, a sudden endoscopic-camera retraction, or a strong illumination change. This paper provides researchers in the medical-imaging computing community with an extensive comparison of the most up-to-date feature-matching algorithms over a large (and annotated) data set of 100 MIS-image pairs obtained from real interventions. The

accuracy

of these methods, as well as their ability to

consistently retrieve

as many good matches as possible, are evaluated for popular feature detectors. In addition, the dataset and the software implementations of these methods are made freely available on the Internet.

In this paper we present the first solution to 3D reconstruction in monocular laparoscopy using methods based on Photometric Stereo (PS). Our main contributions are to provide the new theory and practical solutions to successfully apply PS in close-range imaging conditions. We are specifically motivated by a solution with minimal hardware modification to existing laparoscopes. In fact the only physical modification we make is to adjust the colour of the laparoscope’s illumination via three colour filters placed at its tip. Once calibrated, our approach can compute 3D from a single image, does not require correspondence estimation, and computes absolute depth densely. We demonstrate the potential of our approach with ground truth ex-vivo and in-vivo experimentation.

Image registration is an important tool for imaging validation studies investigating the effect of underlying focal disease on the imaging signal. The strength of the conclusions drawn from these analyses is limited by statistical power. Based on the observation that in this context, statistical power depends in part on uncertainty arising from registration error, we derive a power calculation formula relating registration error, sample size, and the minimum detectable difference between normal and pathologic regions on imaging. Statistical mappings between target registration error and fractional overlap metrics are also derived, and Monte Carlo simulations are used to evaluate the derived models and test the strength of their assumptions.

In this paper we propose a novel graph-based concurrent registration and segmentation framework. Registration is modeled with a pairwise graphical model formulation that is modular with respect to the data and regularization term. Segmentation is addressed by adopting a similar graphical model, using image-based classification techniques while producing a smooth solution. The two problems are coupled via a relaxation of the registration criterion in the presence of tumors as well as a segmentation through a registration term aiming the separation between healthy and diseased tissues. Efficient linear programming is used to solve both problems simultaneously. State of the art results demonstrate the potential of our method on a large and challenging low-grade glioma data set.

Non-rigid image registration using free-form deformations (FFD) is a widely used technique in medical image registration. The balance between robustness and accuracy is controlled by the control point grid spacing and the amount of regularization. In this paper, we revisit the classic FFD registration approach and propose a sparse representation for FFDs using the principles of compressed sensing. The sparse free-form deformation model (SFFD) can capture fine local details such as motion discontinuities without sacrificing robustness. We demonstrate the capabilities of the proposed framework to accurately estimate smooth as well as discontinuous deformations in 2D and 3D image sequences. Compared to the classic FFD approach, a significant increase in registration accuracy can be observed in natural images (61%) as well as in cardiac MR images (53%) with discontinuous motions.

We propose a novel method for registration of 3D fetal brain ultrasound and a reconstructed magnetic resonance fetal brain volumes. The reconstructed MR volume is first segmented using a probabilistic atlas and an ultrasound-like image volume is simulated from the segmentation of the MR image. This ultrasound-like image volume is then affinely aligned with real ultrasound volumes of 27 fetal brains using a robust block-matching approach which can deal with intensity artefacts and missing features in ultrasound images. We show that this approach results in good overlap of four small structures. The average of the co-aligned US images shows good correlation with anatomy of the fetal brain as seen in the MR reconstruction.