Abstract

Segmentation and analysis of sub-cortical structures is of interest in diagnosing some neurological diseases. Segmentation is a challenging task because of brain tissue ambiguity and data scarcity. Deep learning (DL) solutions are widely used for this purpose by considering the problem as a semantic segmentation of the brain. In general, DL approaches exhibit a bias towards larger structures when training is done on the whole brain. We propose a method to address this problem wherein a pre-training step is used to learn tissue characteristics and a rough ROI extraction step aids focusing on local context. We use a Residual U-net for demonstrating the proposed method. Experiments on the IBSR and MICCAI datasets show that our proposed solution leads to an improvement in segmentation performance in general with medium and small size structures benefiting significantly. The performance with the proposed method is also marginally better than a more complex, state of art sub-cortical structure segmentation method. A strength of the proposed method is that it can also be applied as a modification to any existing segmentation solution.

Abstract

Anatomical variabilities seen in longitudinal data or inter- subject data is usually described by the underlying deformation, cap- tured by non-rigid registration of these images. Stationary Velocity Field (SVF) based non-rigid registration algorithms are widely used for reg- istration. However, these methods cover only a limited degree of defor- mations. We address this limitation and define an approximate metric space for the manifold of diffeomorphisms G. We propose a method to break down the large deformation into finite set of small sequential defor- mations. This results in a broken geodesic path on G and its length now forms an approximate registration metric. We illustrate the method using a simple, intensity-based, log-demon implementation. Validation results of the proposed method show that it can capture large and complex de- formations while producing qualitatively better results than state-of-the- art methods. The results also demonstrate that the proposed registration metric is a good indicator of the degree of deformation.

Abstract

Deep learning based methods have shown great promise in achieving accurate automatic detection of Coronavirus Disease (COVID) - 19 from Chest X-Ray ( CXR) images.However, incorporating explainability in these so- lutions remains relatively less explored. We present a hi- erarchical classification approach for separating normal, non- COVID pneumonia (NCP ) and COVID cases using CXR im- ages. We demonstrate that the proposed method achieves clinically consistent explainations. We achieve this using a novel multi-scale attention architecture called Multi-scale Attention Residual Learning (MARL ) and a new loss function based on conicity for training the proposed architecture. The proposed classification strategy has two stages. The first stage uses a model derived from DenseNet to sepa- rate pneumonia cases from normal cases while the second stage uses the MARL architecture to discriminate between COVID and NCP cases. With a five-fold cross validation the proposed method achieves 93%, 96.28%, and 84.51% accu- racy respectively over three large, public datasets for nor- mal vs. NCP vs. COVID classification. This is competitive to the state-of-the-art methods. We also provide explanations in the form of GradCAM attributions, which are well aligned with expert annotations. The attributions are also seen to clearly indicate that MARL deems the peripheral regions of the lungs to be more important in the case of COVID cases while central regions are seen as more important in NCP cases. This observation matches the criteria described by radiologists in clinical literature, thereby attesting to the utility of the derived explanations.

Abstract

Alzheimer’s disease (AD) and Mild Cognitive Impairment (MCI) are neurogenerative impairments with similar symptoms and risk factors. Sulcal width and depth are known biomarkers for discriminating between AD and MCI. This paper presents a novel 2D image representation for a brain mesh surface, called a height map. The basic idea behind the height map is to represent the surface as a function of spherical coordinates of the mesh vertices. We present a method to derive a height map from a given neuroimage (MRI) and extract sulcal regions from the height map. We demonstrate the height map’s utility for classifying a given neuroimage into healthy, MCI and AD classes. Two approaches for extracting sulcal regions are explored. The proposed method is computationally light, and obtaining sulcal regions from a brain surface mesh takes about 24 seconds on a standard Intel i5-7200 CPU. The proposed method achieves 76.1% accuracy, and 76.3% F1-score for healthy, MCI, AD classification on a publicly available dataset.

Abstract

Self-supervised learning is emerging as an effective substitute for transfer learning from large datasets. In this work, we use kidney segmentation to explore this idea. The anatomical asymmetry of kidneys is leveraged to define an effective proxy task for kidney segmentation via self-supervised learning. A siamese convolutional neural network (CNN) is used to classify a given pair of kidney sections from CT volumes as being kidneys of the same or different sides. This knowledge is then transferred for the segmentation of kidneys using another deep CNN using one branch of the siamese CNN as the encoder for the segmentation network. Evaluation results on a publicly available dataset containing computed tomography (CT) scans of the abdominal region shows that a boost in performance and fast convergence can be had relative to a network trained conventionally from scratch. This is notable given that no additional data/expensive annotations or augmentation were used in training.

Abstract

Early detection and treatment of glaucoma is of interest as it is a chronic eye disease leading to an irreversible loss of vision. Existing automated systems rely largely on fundus images for assessment of glaucoma due to their fast acquisition and costeffectiveness. Optical Coherence Tomographic (OCT) images provide vital and unambiguous information about nerve fiber loss and optic cup morphology, which are essential for disease assessment. However, the high cost of OCT is a deterrent for deployment in screening at large scale. In this article, we present a novel CAD solution wherein both OCT and fundus modality images are leveraged to learn a model that can perform a mapping of fundus to OCT feature space. We show how this model can be subsequently used to detect glaucoma given an image from only one modality (fundus). The proposed model has been validated extensively on four public andtwo private datasets. It attained an AUC/Sensitivity value of 0.9429/0.9044 on a diverse set of 568 images, which is superior to the figures obtained by a model that is trained only on fundus features. Cross-validation was also done on nearly 1,600 images drawn from a private (OD-centric) and a public (macula-centric) dataset and the proposed model was found to outperform the state-of-the-art method by 8% (public) to 18% (private). Thus, we conclude that fundus to OCT feature space mapping is an attractive option for glaucoma detection.

Abstract

Segmentation of sub-cortical structures from MRI scans is of interest in many neurological diagnosis. Since this is a laborious task machine learning and specifically deep learning (DL) methods have become explored. The structural complexity of the brain demands a large, high quality segmentation dataset to develop good DL-based solutions for sub-cortical structure segmentation. Towards this, we are releasing a set of 114, 1.5 Tesla, T1 MRI scans with manual delineations for 14 sub-cortical structures. The scans in the dataset were acquired from healthy young (21-30 years) subjects ( 58 male and 56 female) and all the structures are manually delineated by experienced radiology experts. Segmentation experiments have been conducted with this dataset and results demonstrate that accurate results can be obtained with deep-learning methods.

Abstract

A major challenge that deep learning systems face is the Catastrophic Forgetting (CF) phenomenon that is observed when finetuning is used to try and adapt a system to a new task or a sequence of datasets with different distributions. CF refers to the significant degradation in performance on the old task/dataset. In this paper, a novel approach is proposed to address CF in computer aided diagnosis (CAD) system design in the medical domain. CAD systems often need to handle a sequence of datasets collected over time from different sites with different imaging parameters/populations. The solution we propose is to move samples from all the datasets closer to a common manifold via a reformer at the front end of a CAD system. The utility of this approach is demonstrated on two common tasks, namely segmentation and classification, using publicly available datasets. Results of extensive experiments show that manifold learning can yield about 74% improvement on an average in the reduction of CF over the baseline fine-tuning process and the state-of-the-art regularization based methods. The results also indicate that a Reformer when used in conjunction with the state-of-the-art regularization methods, has the potential to yield further improvement in CF reduction.

Abstract

Explainable AI aims to build systems that not only give high performance but also are able to provide insights that drive the decision making. However, deriving this explanation is often dependent on fully annotated (class label and local annotation) data, which are not readily available in the medical domain. Approach: This paper addresses the above-mentioned aspects and presents an innovative approach to classifying a lung nodule in a CT volume as malignant or benign, and generating a morphologically meaningful explanation for the decision in the form of attributes such as nodule margin, sphericity, and spiculation. A deep learning architecture that is trained using a multi-phase training regime is proposed. The nodule class label (benign/malignant) is learned with full supervision and is guided by semantic attributes that are learned in a weakly supervised manner. Results: Results of an extensive evaluation of the proposed system on the LIDC-IDRI dataset show good performance compared with state-of-the-art, fully supervised methods. The proposed model is able to label nodules (after full supervision) with an accuracy of 89.1% and an area under curve of 0.91 and to provide eight attributes scores as an explanation, which is learned from a much smaller training set. The proposed system's potential to be integrated with a sub-optimal nodule detection system was also tested, and our system handled 95% of false positive or random regions in the input well by labeling them as benign, which underscores its robustness. Conclusions: The proposed approach offers a way to address computer-aided diagnosis system design under the constraint of sparse availability of fully annotated images.

Abstract

This work explores a hybrid approach to segmentation as an alternative to a purely data-driven approach. We introduce an end-to-end U-Net based network called DU-Net, which uses additional frequency preserving features, namely the Scattering Coefficients (SC), for medical image segmentation. SC are translation invariant and Lipschitz continuous to deformations which help DU-Net outperform other conventional CNN counterparts on four datasets and two segmentation tasks: Optic Disc and Optic Cup in color fundus images and fetal Head in ultrasound images. The proposed method shows remarkable improvement over the basic U-Net with performance competitive to state-of-the-art methods. The results indicate that it is possible to use a lighter network trained with fewer images (without any augmentation) to attain good segmentation results.