Deep Medical Image Analysis From Low-Resource Shape Coding

Medical image analysis is crucial in modern healthcare, enabling the quantification and identification of anatomical structures to characterize disease processes, assess treatment effects, and personalize patient healthcare. However, conventional medical image analysis frameworks require substantial resources, including high-quality imaging, extensive expert annotations, and scalable computational power. These demands limit their reliability in low resource settings, where access to such resources is constrained. Addressing these challenges is essential to ensure that advanced medical imaging techniques can benefit diverse healthcare environments, ultimately improving global access to high-quality care. This thesis aims to develop a self-integrated framework to overcome the key limitations f low-resource medical image analysis. We categorize low-resource medical image analysis into three primary research areas: (1) Feature extraction from low-quality medical images, where we introduce a novel token-matching framework to mitigate pixel-level anatomical ambiguities. (2) Multi-task learning with limited annotations, where we propose a semisupervised framework to establish task correspondences with few-shot examples despite the scarcity of high-quality labels. (3) Concept discovery with lightweight computation, where we develop a prompt-learning framework that eliminates the need for extensive training or fine-tuning configurations. Our proposed techniques enable a range of downstream applications, including biomarker segmentation in echocardiograph images, deformable landmark tracking in ultrasound scans, and vision-language segmentation through visual prompt learning. By integrating deep learning frameworks with low-resource shape coding, this thesis advances accurate, robust, and highly personalized decision-making in medical imaging.