This thesis investigates how domain knowledge can be incorporated into medical image analysis, synthesis, and prediction to improve accessibility and clinical reliability under realistic resource constraints. Although deep learning has achieved strong performance in controlled medical imaging benchmarks, practical deployment often remains limited by degraded image quality, scarce or imperfect supervision, constrained access to advanced imaging systems, and the mismatch between standard optimization objectives and clinically meaningful validity. To address these challenges, this thesis develops two complementary research directions: robust medical image analysis under low-quality and sparse annotation, and clinically informed medical image synthesis and prediction under limited paired data and constrained imaging resources. The first part of the thesis focuses on robust analysis under degraded inputs and constrained supervision. A similarity-weighted self-ensembling framework is developed for low-quality few-shot stroke lesion segmentation, improving robustness by explicitly addressing both image degradation and annotation scarcity. The second part investigates clinically meaningful synthesis and prediction through self-supervision, anatomical priors, and semantic guidance. A self-supervised anatomical continuity enhancement network is proposed for generating 7T susceptibility-weighted imaging from 3T inputs under limited paired data. A self-supervised T2WI-bridged framework is further developed for joint liver segmentation and proton density fat fraction prediction from ultrasound, reducing dependence on MRI-based quantification. Finally, an anatomy-grounded weak clinical supervision framework is proposed for pathology-preserving low-field brain MRI enhancement, using diagnostic reports during training to introduce clinically grounded semantic priors without requiring report input at inference time. Taken together, these studies show that medical image learning becomes more practically useful when robustness, domain knowledge, and clinical meaning are addressed jointly rather than in isolation. This thesis contributes a principled framework for accessible medical imaging in which anatomically valid, pathology-relevant, and clinically trustworthy outputs can be achieved even when data quality, supervision, and imaging resources are constrained.
Dong Zhang (Fri,) studied this question.