Learning Data Representations for Improving Detection Performance of Deep Learning Models in Medical Imaging Applications

Abstract

Supervised deep learning (DL) models, with large, labeled datasets, are currently thestate-of-the-art in several classification and segmentation tasks in medical imaging applications. However, obtaining expert manual annotations for data hungry DL models is time-consuming. Additionally, the performance of these algorithms can be limited by the choice of data representation on which they are applied. A good data representation can make the subsequent classification task easier. The goal of this dissertation is to learn data representations that can improve the performance of DL models for medical image segmentation tasks, especially Magnetic Resonance (MR) images, and leverage limited availability of labeled data. In this work, we propose supervised and self-supervised strategies to transform inputimages to its new representation and improve the classification performance of DL models. For supervised feature engineering of such representations, we propose 1) using a stacked generalization ensemble of models to learn new data representation in predictions’ space where subsequent classification is performed and 2) applying domain knowledge to transform images to a new MR contrast where classification is performed in the synthesized contrast space. Motivated by the abundance of unlabeled MR images compared to labeled ones, we incorporate domain knowledge with representational learning to learn efficient representations of MR images in feature space. Specifically, we propose a novel contrastive learning strategy, that utilizes tissue-specific constraints derived from multi-contrast MR images, to provide a suitable initialization for data hungry DL models. We also investigate the improvements in model performance compared to conventional initialization techniques. The utility of each of the proposed strategies is evaluated in different MR image segmentationtasks. These include a) detecting white matter hyperintensities in T2-weighted MR images of the brain for lesion burden assessment, b) detecting thalamus and its sub-thalamic nuclei in T1-weighted MR images of the brain for quantifying age-related atrophy, and c) multi-organ segmentation in MR images.