Few-Shot Medical Image Segmentation Using Semi-Supervised and Weakly-Supervised Learning

Abstract

Medical image segmentation plays a critical role in assessing, diagnosing, and treating various medical conditions. State-of-the-art convolutional neural network architectures have demonstrated promising performance. However, these methods heavily rely on labeled images to achieve optimal performance, which is challenging, particularly in medical cases where obtaining pixel-level annotations is often difficult. This thesis explores six distinct methods for the few-shot segmentation of medical images, with a specific focus on cardiac structures: 1- We propose a novel weakly- supervised learning method for the rapid annotation of the left ventricle and based on its circular shape. 2-We introduce a novel physics-informed label-propagation method based on the patient-specific characteristics of the left ventricle to transform the labels from end-diastole and end-systole phases through the entire cardiac cycle. 3- We combine this label propagation method with a semi-supervised learning approach to provide a labeling framework for the segmentation of the left ventricle with severely limited labeled data. 4- We propose a left ventricle segmentation method inspired by the idea of consistency regularization and by creating redundant complementing labels. 5- We introduce a semi-supervised learning method for the segmentation of any medical image by designing a novel loss function. The loss function incorporates an adaptation of active contours and also considers reliable and unreliable pixels through masked cross-entropy and masked active contour terms. 6- We develop a framework for enhancing medical image segmentation by using different techniques, including total variation, deformable models, and uncertainty-based pixel-level and image-level pseudo-label pruning. We evaluated our proposed methods on various medical image datasets, including SCD, ACDC, and CCLISD. These datasets segment different structures from CT and MRI modalities. Our methods significantly enhance the segmentation accuracy in terms of Dice index and Jaccard index while also reducing the required resources, including the time for each annotation, expertise, and the number of annotations.