Novel Machine Learning Techniques for Left Ventricular Analysis in Echocardiography

Abstract

Transthoracic (TTX) echocardiography (echo) is vital for the diagnosis and treatment of heart disease. It is also essential for the determination of appropriate therapeutic procedures and monitoring disease progression and response. An important requisite of TTX echo is the quantitative assessment of left ventricular (LV) size and function from its manually traced endocardial border. However, the diagnostic accuracy of this routine task is often adversely affected by artifact and signal dropout, producing substantial observer variability and uncertainty in clinical diagnosis. With the advent of machine learning, computer aided detection (CAD) systems have addressed organ segmentation through robust landmark localization techniques and a reliance on anatomical shape prior models to guide the segmentation process. Shape priors models are most effective when shape variations can be captured by a parametric distribution, and suffcient training data is available. However, in the absence of these conditions, results are invariably much poorer. In addition, the problem of insuffcient training data not only presents challenges to shape prior models, but also classification algorithms, as well. Stress echo (SE) is a widely used functional test for the detection of obstructive coronary artery disease. The interpretive process involves the careful comparison of pre- and post-exercise echo sequences across a number of echocardiographic views. This is a time consuming and highly subjective task. While CAD systems have been shown to be feasible for automated reporting in other areas of medical imaging, they have not been applied to the task of identifying abnormal stress echocardiograms. An automated approach would not only be a useful adjunct to physicians reporting SE, but also aid in physician training of SE reporting. Motivated by these challenges, this thesis presents four novel machine learning techniques for LV analysis in echo. Firstly, two anatomical shape prior models are proposed: online relational manifold learning (ORML) and dual subspace segment projection learning (DSSPL). ORML is formulated to address the challenge of modelling complex shape subspaces. ORML serves to learn a mapping function between a low dimension image manifold and shape manifold. However, different to existing subspace learning approaches, ORML leverages the input image to target more contextually relevant regions between both manifold structures, leading to robust LV shape inference for volume prediction, and the formulation of a shape prior model through a more principled shape selection strategy. ORML demonstrates improved segmentation performance over current benchmark methods, and shows an excellent level of agreement with an expert. DSSPL addresses the challenges of modelling complex shape variations under the scenario of high dimension low sample size (HDLSS) training data. It serves to compose shapes from an ensemble of shape segments where each segment is formed using two subspaces: global shape subspace and segment-specific subspace, each necessary for extracting global shape patterns and local patterns, respectively. This ensures general shape plausibility in regions of signal drop-out or missing boundary information, and also more localized flexibility. The reconstructive properties of DSSPL reduces information loss and leverages the subspaces to provide contiguous shapes without any post-processing. Comprehensive experimental analysis is performed on three databases from different medical imaging systems across X-Ray, MRI, and echo. DSSPL outperforms all compared benchmarks in terms of its shape generalization ability and segmentation performance. The third method proposed is dual subspace discriminative projection learning (DSDPL), which addresses the challenge of image classification, also under the HDLSS training data scenario. Unlike traditional projection learning frameworks that assume discriminative features share a common subspace, DSDPL instead serves to decompose original high dimensional data, via learned projection matrices, into class-shared and class-specific subspaces. The learned projection matrices are jointly constrained with l2;1 sparse norm and LDA terms while the reconstructive properties reduce information loss. Regression-based terms are also included to facilitate a more robust classification approach, using extracted class-specific features for better classification. Results show improved classification accuracy with DSDPL over current benchmark subspace learning methods and deep learning models. The fourth method proposed is deep stage-coupled attentive feature extraction (DSCAFE) for identifying abnormal stress echocardiograms. DSCAFE is a deep neural network model that consists of stage-coupled attentive feature extraction (SCAFE) blocks for extracting the most salient information from connected echo sequences. SCAFE blocks are composed of 3D residual network streams and dual-attention gated mechanisms, which provide more targeted focus across each echo sequence by also taking into account the observed features from the corresponding view at the opposing exercise stage. A recurrent neural network feature aggregation strategy is then employed to model the extracted low dimension spatio-temporal features for more accurate classification. When compared against an expert reviewer, DSCAFE achieved a concordance of 86.5% from a clinical SE dataset. This research draws on machine learning knowledge across a diverse range of domains. While results show definitive improvements over current benchmark methods for LV analysis, the proposed methods are also adaptable to a wide range of computer vision tasks.