Brain extraction from normal and pathological images: A joint PCA/Image-Reconstruction approach.

Brain extraction from 3D medical images is a common pre-processing step. A variety of approaches exist, but they are frequently only designed to perform brain extraction from images without strong pathologies. Extracting the brain from images exhibiting strong pathologies, for example, the presence of a brain tumor or of a traumatic brain injury (TBI), is challenging. In such cases, tissue appearance may substantially deviate from normal tissue appearance and hence violates algorithmic assumptions for standard approaches to brain extraction; consequently, the brain may not be correctly extracted. This paper proposes a brain extraction approach which can explicitly account for pathologies by jointly modeling normal tissue appearance and pathologies. Specifically, our model uses a three-part image decomposition: (1) normal tissue appearance is captured by principal component analysis (PCA), (2) pathologies are captured via a total variation term, and (3) the skull and surrounding tissue is captured by a sparsity term. Due to its convexity, the resulting decomposition model allows for efficient optimization. Decomposition and image registration steps are alternated to allow statistical modeling of normal tissue appearance in a fixed atlas coordinate system. As a beneficial side effect, the decomposition model allows for the identification of potentially pathological areas and the reconstruction of a quasi-normal image in atlas space. We demonstrate the effectiveness of our approach on four datasets: the publicly available IBSR and LPBA40 datasets which show normal image appearance, the BRATS dataset containing images with brain tumors, and a dataset containing clinical TBI images. We compare the performance with other popular brain extraction models: ROBEX, BEaST, MASS, BET, BSE and a recently proposed deep learning approach. Our model performs better than these competing approaches on all four datasets. Specifically, our model achieves the best median (97.11) and mean (96.88) Dice scores over all datasets. The two best performing competitors, ROBEX and MASS, achieve scores of 96.23/95.62 and 96.67/94.25 respectively. Hence, our approach is an effective method for high quality brain extraction for a wide variety of images.

Quicksilver: Fast predictive image registration - A deep learning approach.

This paper introduces Quicksilver, a fast deformable image registration method. Quicksilver registration for image-pairs works by patch-wise prediction of a deformation model based directly on image appearance. A deep encoder-decoder network is used as the prediction model. While the prediction strategy is general, we focus on predictions for the Large Deformation Diffeomorphic Metric Mapping (LDDMM) model. Specifically, we predict the momentum-parameterization of LDDMM, which facilitates a patch-wise prediction strategy while maintaining the theoretical properties of LDDMM, such as guaranteed diffeomorphic mappings for sufficiently strong regularization. We also provide a probabilistic version of our prediction network which can be sampled during the testing time to calculate uncertainties in the predicted deformations. Finally, we introduce a new correction network which greatly increases the prediction accuracy of an already existing prediction network. We show experimental results for uni-modal atlas-to-image as well as uni-/multi-modal image-to-image registrations. These experiments demonstrate that our method accurately predicts registrations obtained by numerical optimization, is very fast, achieves state-of-the-art registration results on four standard validation datasets, and can jointly learn an image similarity measure. Quicksilver is freely available as an open-source software.

ICON: Learning Regular Maps Through Inverse Consistency.

Learning maps between data samples is fundamental. Applications range from representation learning, image translation and generative modeling, to the estimation of spatial deformations. Such maps relate feature vectors, or map between feature spaces. Well-behaved maps should be regular, which can be imposed explicitly or may emanate from the data itself. We explore what induces regularity for spatial transformations, e.g., when computing image registrations. Classical optimization-based models compute maps between pairs of samples and rely on an appropriate regularizer for well-posedness. Recent deep learning approaches have attempted to avoid using such regularizers altogether by relying on the sample population instead. We explore if it is possible to obtain spatial regularity using an inverse consistency loss only and elucidate what explains map regularity in such a context. We find that deep networks combined with an inverse consistency loss and randomized off-grid interpolation yield well behaved, approximately diffeomorphic, spatial transformations. Despite the simplicity of this approach, our experiments present compelling evidence, on both synthetic and real data, that regular maps can be obtained without carefully tuned explicit regularizers, while achieving competitive registration performance.

Comparing Object Recognition in Humans and Deep Convolutional Neural Networks-An Eye Tracking Study.

Deep convolutional neural networks (DCNNs) and the ventral visual pathway share vast architectural and functional similarities in visual challenges such as object recognition. Recent insights have demonstrated that both hierarchical cascades can be compared in terms of both exerted behavior and underlying activation. However, these approaches ignore key differences in spatial priorities of information processing. In this proof-of-concept study, we demonstrate a comparison of human observers (N = 45) and three feedforward DCNNs through eye tracking and saliency maps. The results reveal fundamentally different resolutions in both visualization methods that need to be considered for an insightful comparison. Moreover, we provide evidence that a DCNN with biologically plausible receptive field sizes called vNet reveals higher agreement with human viewing behavior as contrasted with a standard ResNet architecture. We find that image-specific factors such as category, animacy, arousal, and valence have a direct link to the agreement of spatial object recognition priorities in humans and DCNNs, while other measures such as difficulty and general image properties do not. With this approach, we try to open up new perspectives at the intersection of biological and computer vision research.

An empirical assessment of appearance descriptors applied to MRI for automated diagnosis of TLE and MCI.

OBJECTIVE: Differential diagnosis of mild cognitive impairment MCI and temporal lobe epilepsy TLE is a debated issue, specifically because these conditions may coincide in the elderly population. We evaluate automated differential diagnosis based on characteristics derived from structural brain MRI of different brain regions. METHODS: In 22 healthy controls, 19 patients with MCI, and 17 patients with TLE we used scale invariant feature transform (SIFT), local binary patterns (LBP), and wavelet-based features and investigate their predictive performance for MCI and TLE. RESULTS: The classification based on SIFT features resulted in an accuracy of 81% of MCI vs. TLE and reasonable generalizability. Local binary patterns yielded satisfactory diagnostic performance with up to 94.74% sensitivity and 88.24% specificity in the right Thalamus for the distinction of MCI vs. TLE, but with limited generalizable. Wavelet features yielded similar results as LPB with 94.74% sensitivity and 82.35% specificity but generalize better. SIGNIFICANCE: Features beyond volume analysis are a valid approach when applied to specific regions of the brain. Most significant information could be extracted from the thalamus, frontal gyri, and temporal regions, among others. These results suggest that analysis of changes of the central nervous system should not be limited to the most typical regions of interest such as the hippocampus and parahippocampal areas. Region-independent approaches can add considerable information for diagnosis. We emphasize the need to characterize generalizability in future studies, as our results demonstrate that not doing so can lead to overestimation of classification results. LIMITATIONS: The data used within this study allows for separation of MCI and TLE subjects using a simple age threshold. While we present a strong indication that the presented method is age-invariant and therefore agnostic to this situation, new data would be needed for a rigorous empirical assessment of this findings.

Metric Learning for Image Registration.

Image registration is a key technique in medical image analysis to estimate deformations between image pairs. A good deformation model is important for high-quality estimates. However, most existing approaches use ad-hoc deformation models chosen for mathematical convenience rather than to capture observed data variation. Recent deep learning approaches learn deformation models directly from data. However, they provide limited control over the spatial regularity of transformations. Instead of learning the entire registration approach, we learn a spatially-adaptive regularizer within a registration model. This allows controlling the desired level of regularity and preserving structural properties of a registration model. For example, diffeomorphic transformations can be attained. Our approach is a radical departure from existing deep learning approaches to image registration by embedding a deep learning model in an optimization-based registration algorithm to parameterize and data-adapt the registration model itself. Source code is publicly-available at https://github.com/uncbiag/registration.

Fast predictive simple geodesic regression.

Deformable image registration and regression are important tasks in medical image analysis. However, they are computationally expensive, especially when analyzing large-scale datasets that contain thousands of images. Hence, cluster computing is typically used, making the approaches dependent on such computational infrastructure. Even larger computational resources are required as study sizes increase. This limits the use of deformable image registration and regression for clinical applications and as component algorithms for other image analysis approaches. We therefore propose using a fast predictive approach to perform image registrations. In particular, we employ these fast registration predictions to approximate a simplified geodesic regression model to capture longitudinal brain changes. The resulting method is orders of magnitude faster than the standard optimization-based regression model and hence facilitates large-scale analysis on a single graphics processing unit (GPU). We evaluate our results on 3D brain magnetic resonance images (MRI) from the ADNI datasets.

Patient-Specific Registration of Pre-operative and Post-recurrence Brain Tumor MRI Scans.

Registering brain magnetic resonance imaging (MRI) scans containing pathologies is challenging primarily due to large deformations caused by the pathologies, leading to missing correspondences between scans. However, the registration task is important and directly related to personalized medicine, as registering between baseline pre-operative and post-recurrence scans may allow the evaluation of tumor infiltration and recurrence. While many registration methods exist, most of them do not specifically account for pathologies. Here, we propose a framework for the registration of longitudinal image-pairs of individual patients diagnosed with glioblastoma. Specifically, we present a combined image registration/reconstruction approach, which makes use of a patient-specific principal component analysis (PCA) model of image appearance to register baseline pre-operative and post-recurrence brain tumor scans. Our approach uses the post-recurrence scan to construct a patient-specific model, which then guides the registration of the pre-operative scan. Quantitative and qualitative evaluations of our framework on 10 patient image-pairs indicate that it provides excellent registration performance without requiring (1) any human intervention or (2) prior knowledge of tumor location, growth or appearance.

Image-Based Methods for Phase Estimation, Gating, and Temporal Superresolution of Cardiac Ultrasound.

OBJECTIVE: Ultrasound is an effective tool for rapid noninvasive assessment of cardiac structure and function. Determining the cardiorespiratory phases of each frame in the ultrasound video and capturing the cardiac function at a much higher temporal resolution are essential in many applications. Fulfilling these requirements is particularly challenging in preclinical studies involving small animals with high cardiorespiratory rates, requiring cumbersome and expensive specialized hardware. METHODS: We present a novel method for the retrospective estimation of cardiorespiratory phases directly from the ultrasound videos. It transforms the videos into a univariate time series preserving the evidence of periodic cardiorespiratory motion, decouples the signatures of cardiorespiratory motion with a trend extraction technique, and estimates the cardiorespiratory phases using a Hilbert transform approach. We also present a robust nonparametric regression technique for respiratory gating and a novel kernel-regression model for reconstructing images at any cardiac phase facilitating temporal superresolution. RESULTS: We validated our methods using two-dimensional echocardiography videos and electrocardiogram (ECG) recordings of six mice. Our cardiac phase estimation method provides accurate phase estimates with a mean-phase-error range of 3%-6% against ECG derived phase and outperforms three previously published methods in locating ECGs R-wave peak frames with a mean-frame-error range of 0.73-1.36. Our kernel-regression model accurately reconstructs images at any cardiac phase with a mean-normalized-correlation range of 0.81-0.85 over 50 leave-one-out-cross-validation rounds. CONCLUSION AND SIGNIFICANCE: Our methods can enable tracking of cardiorespiratory phases without additional hardware and reconstruction of respiration-free single cardiac-cycle videos at a much higher temporal resolution.

Quest for the best endoscopic imaging modality for computer-assisted colonic polyp staging.

BACKGROUND: It was shown in previous studies that high definition endoscopy, high magnification endoscopy and image enhancement technologies, such as chromoendoscopy and digital chromoendoscopy [narrow-band imaging (NBI), i-Scan] facilitate the detection and classification of colonic polyps during endoscopic sessions. However, there are no comprehensive studies so far that analyze which endoscopic imaging modalities facilitate the automated classification of colonic polyps. In this work, we investigate the impact of endoscopic imaging modalities on the results of computer-assisted diagnosis systems for colonic polyp staging. AIM: To assess which endoscopic imaging modalities are best suited for the computer-assisted staging of colonic polyps. METHODS: In our experiments, we apply twelve state-of-the-art feature extraction methods for the classification of colonic polyps to five endoscopic image databases of colonic lesions. For this purpose, we employ a specifically designed experimental setup to avoid biases in the outcomes caused by differing numbers of images per image database. The image databases were obtained using different imaging modalities. Two databases were obtained by high-definition endoscopy in combination with i-Scan technology (one with chromoendoscopy and one without chromoendoscopy). Three databases were obtained by high-magnification endoscopy (two databases using narrow band imaging and one using chromoendoscopy). The lesions are categorized into non-neoplastic and neoplastic according to the histological diagnosis. RESULTS: Generally, it is feature-dependent which imaging modalities achieve high results and which do not. For the high-definition image databases, we achieved overall classification rates of up to 79.2% with chromoendoscopy and 88.9% without chromoendoscopy. In the case of the database obtained by high-magnification chromoendoscopy, the classification rates were up to 81.4%. For the combination of high-magnification endoscopy with NBI, results of up to 97.4% for one database and up to 84% for the other were achieved. Non-neoplastic lesions were classified more accurately in general than non-neoplastic lesions. It was shown that the image recording conditions highly affect the performance of automated diagnosis systems and partly contribute to a stronger effect on the staging results than the used imaging modality. CONCLUSION: Chromoendoscopy has a negative impact on the results of the methods. NBI is better suited than chromoendoscopy. High-definition and high-magnification endoscopy are equally suited.

Compressing Networks with Super Nodes.

Community detection is a commonly used technique for identifying groups in a network based on similarities in connectivity patterns. To facilitate community detection in large networks, we recast the network as a smaller network of 'super nodes', where each super node comprises one or more nodes of the original network. We can then use this super node representation as the input into standard community detection algorithms. To define the seeds, or centers, of our super nodes, we apply the 'CoreHD' ranking, a technique applied in network dismantling and decycling problems. We test our approach through the analysis of two common methods for community detection: modularity maximization with the Louvain algorithm and maximum likelihood optimization for fitting a stochastic block model. Our results highlight that applying community detection to the compressed network of super nodes is significantly faster while successfully producing partitions that are more aligned with the local network connectivity and more stable across multiple (stochastic) runs within and between community detection algorithms, yet still overlap well with the results obtained using the full network.

Training of polyp staging systems using mixed imaging modalities.

BACKGROUND: In medical image data sets, the number of images is usually quite small. The small number of training samples does not allow to properly train classifiers which leads to massive overfitting to the training data. In this work, we investigate whether increasing the number of training samples by merging datasets from different imaging modalities can be effectively applied to improve predictive performance. Further, we investigate if the extracted features from the employed image representations differ between different imaging modalities and if domain adaption helps to overcome these differences. METHOD: We employ twelve feature extraction methods to differentiate between non-neoplastic and neoplastic lesions. Experiments are performed using four different classifier training strategies, each with a different combination of training data. The specifically designed setup for these experiments enables a fair comparison between the four training strategies. RESULTS: Combining high definition with high magnification training data and chromoscopic with non-chromoscopic training data partly improved the results. The usage of domain adaptation has only a small effect on the results compared to just using non-adapted training data. CONCLUSION: Merging datasets from different imaging modalities turned out to be partially beneficial for the case of combining high definition endoscopic data with high magnification endoscopic data and for combining chromoscopic with non-chromoscopic data. NBI and chromoendoscopy on the other hand are mostly too different with respect to the extracted features to combine images of these two modalities for classifier training.

SCOLIOSIS SCREENING AND MONITORING USING SELF CONTAINED ULTRASOUND AND NEURAL NETWORKS.

We aim to diagnose scoliosis using a self contained ultrasound device that does not require significant training to operate. The device knows its angle relative to vertical using an embedded inertial measurement unit, and it estimates its angle relative to a vertebrae using a neural network analysis of its ultrasound images. The composition of those angles defines the angle of a vertebrae from vertical. The maximum difference between vertebrae angles collected from a scan of a spine yields the Cobb angle measure that is used to quantify scoliosis severity.

EFFICIENT REGISTRATION OF PATHOLOGICAL IMAGES: A JOINT PCA/IMAGE-RECONSTRUCTION APPROACH.

Registration involving one or more images containing pathologies is challenging, as standard image similarity measures and spatial transforms cannot account for common changes due to pathologies. Low-rank/Sparse (LRS) decomposition removes pathologies prior to registration; however, LRS is memory-demanding and slow, which limits its use on larger data sets. Additionally, LRS blurs normal tissue regions, which may degrade registration performance. This paper proposes an efficient alternative to LRS: (1) normal tissue appearance is captured by principal component analysis (PCA) and (2) blurring is avoided by an integrated model for pathology removal and image reconstruction. Results on synthetic and BRATS 2015 data demonstrate its utility.

Parametric Regression on the Grassmannian.

We address the problem of fitting parametric curves on the Grassmann manifold for the purpose of intrinsic parametric regression. We start from the energy minimization formulation of linear least-squares in Euclidean space and generalize this concept to general nonflat Riemannian manifolds, following an optimal-control point of view. We then specialize this idea to the Grassmann manifold and demonstrate that it yields a simple, extensible and easy-to-implement solution to the parametric regression problem. In fact, it allows us to extend the basic geodesic model to (1) a "time-warped" variant and (2) cubic splines. We demonstrate the utility of the proposed solution on different vision problems, such as shape regression as a function of age, traffic-speed estimation and crowd-counting from surveillance video clips. Most notably, these problems can be conveniently solved within the same framework without any specifically-tailored steps along the processing pipeline.

Registration of Pathological Images.

This paper proposes an approach to improve atlas-to-image registration accuracy with large pathologies. Instead of directly registering an atlas to a pathological image, the method learns a mapping from the pathological image to a quasi-normal image, for which more accurate registration is possible. Specifically, the method uses a deep variational convolutional encoder-decoder network to learn the mapping. Furthermore, the method estimates local mapping uncertainty through network inference statistics and uses those estimates to down-weight the image registration similarity measure in areas of high uncertainty. The performance of the method is quantified using synthetic brain tumor images and images from the brain tumor segmentation challenge (BRATS 2015).

Group Testing for Longitudinal Data.

We consider how to test for group differences of shapes given longitudinal data. In particular, we are interested in differences of longitudinal models of each group's subjects. We introduce a generalization of principal geodesic analysis to the tangent bundle of a shape space. This allows the estimation of the variance and principal directions of the distribution of trajectories that summarize shape variations within the longitudinal data. Each trajectory is parameterized as a point in the tangent bundle. To study statistical differences in two distributions of trajectories, we generalize the Bhattacharyya distance in Euclidean space to the tangent bundle. This not only allows to take second-order statistics into account, but also serves as our test-statistic during permutation testing. Our method is validated on both synthetic and real data, and the experimental results indicate improved statistical power in identifying group differences. In fact, our study sheds new light on group differences in longitudinal corpus callosum shapes of subjects with dementia versus normal controls.

Low-Rank Atlas Image Analyses in the Presence of Pathologies.

We present a common framework, for registering images to an atlas and for forming an unbiased atlas, that tolerates the presence of pathologies such as tumors and traumatic brain injury lesions. This common framework is particularly useful when a sufficient number of protocol-matched scans from healthy subjects cannot be easily acquired for atlas formation and when the pathologies in a patient cause large appearance changes. Our framework combines a low-rank-plus-sparse image decomposition technique with an iterative, diffeomorphic, group-wise image registration method. At each iteration of image registration, the decomposition technique estimates a "healthy" version of each image as its low-rank component and estimates the pathologies in each image as its sparse component. The healthy version of each image is used for the next iteration of image registration. The low-rank and sparse estimates are refined as the image registrations iteratively improve. For unbiased atlas formation, at each iteration, the average of the low-rank images from the patients is used as the atlas image for the next iteration, until convergence. Since each iteration's atlas is comprised of low-rank components, it provides a population-consistent, pathology-free appearance. Evaluations of the proposed methodology are presented using synthetic data as well as simulated and clinical tumor MRI images from the brain tumor segmentation (BRATS) challenge from MICCAI 2012.

Time-warped geodesic regression.

We consider geodesic regression with parametric time-warps. This allows for example, to capture saturation effects as typically observed during brain development or degeneration. While highly-flexible models to analyze time-varying image and shape data based on generalizations of splines and polynomials have been proposed recently, they come at the cost of substantially more complex inference. Our focus in this paper is therefore to keep the model and its inference as simple as possible while allowing to capture expected biological variation. We demonstrate that by augmenting geodesic regression with parametric time-warp functions, we can achieve comparable flexibility to more complex models while retaining model simplicity. In addition, the time-warp parameters provide useful information of underlying anatomical changes as demonstrated for the analysis of corpora callosa and rat calvariae. We exemplify our strategy for shape regression on the Grassmann manifold, but note that the method is generally applicable for time-warped geodesic regression.

Low-rank to the rescue - atlas-based analyses in the presence of pathologies.

Low-rank image decomposition has the potential to address a broad range of challenges that routinely occur in clinical practice. Its novelty and utility in the context of atlas-based analysis stems from its ability to handle images containing large pathologies and large deformations. Potential applications include atlas-based tissue segmentation and unbiased atlas building from data containing pathologies. In this paper we present atlas-based tissue segmentation of MRI from patients with large pathologies. Specifically, a healthy brain atlas is registered with the low-rank components from the input MRIs, the low-rank components are then re-computed based on those registrations, and the process is then iteratively repeated. Preliminary evaluations are conducted using the brain tumor segmentation challenge data (BRATS '12).