Toward robust and fair ultrasound classification with clinical context

Deep learning has achieved remarkable progress in medical image analysis, yet its clinical translation remains limited by three persistent challenges: imperfect supervision, lack of clinical contextualization, and poor generalization under subpopulation shift. This thesis addresses these challenges through a unified research program focused on developing reliable and fair artificial intelligence for medical imaging. First, we investigate learning under noisy or limited labels, a pervasive issue in ultrasound and other modalities where annotations are coarse, subjective, or expensive to obtain. We introduce label-robust frameworks that refine coarse involvement labels, identify corrupted samples through loss modeling and peer learning, and enhance representation stability through self-supervised pretraining and prototype-based architectures. These methods improve diagnostic accuracy and robustness in prostate ultrasound cancer detection. Second, we address the absence of patient-specific contextual reasoning in current imaging models. We adapt medical foundation models to ultrasound and design multimodal frameworks that jointly process imaging and clinical metadata. The proposed systems, ProstNFound and TREAT-Net, demonstrate that integrating structured clinical variables with visual features improves both generalization and interpretability in prostate cancer detection and echocardiographic treatment prediction. Finally, we examine fairness and robustness under subpopulation shift, where performance disparities emerge across demographic, institutional, or acquisition subgroups. We develop diversified prototypical ensembles (DPE) that explore complementary decision subspaces, and a transformer-based aggregation mechanism (DPE-Former) that adaptively reweights ensemble predictions to reduce the influence of confounded or redundant features. These approaches yield improved worst-group accuracy and stability across multiple medical and benchmark datasets. This thesis demonstrates that building reliable AI for healthcare involves more than achieving high accuracy. It requires developing methods that can handle imperfect training labels, incorporate relevant clinical information, and remain stable when applied to real-world data from diverse settings. The proposed approaches are evaluated in two clinically important domains: prostate ultrasound for cancer detection and echocardiography for the management of acute coronary syndromes. Together, these studies provide both technical insights and practical advances toward creating medical imaging systems that are robust, fair, and informed by clinical context.