Artificial intelligence for hepatopancreatic oncologic imaging Self-supervised learning and clinical translation in cancer assessment

Artificial intelligence (AI) holds clear promise for improving cancer diagnosis and treatment on medical imaging, yet its translation into routine clinical practice remains limited. This thesis examines the gap between AI capability and clinical implementation in hepato-pancreatic oncology, arguing that successful translation requires the coordinated advancement of technical performance, clinical utility, and reliable deployment rather than any one of these in isolation.
The work is organized in three parts. Part I develops the technical foundation: COALA, a self-learning teacher–student model for segmenting colorectal liver metastases, and a tripartite teacher–professor–student framework for locally advanced pancreatic cancer, both achieving strong segmentation accuracy across internal and external CT datasets. Part II translates these segmentation outputs into decision support: VasQNet quantifies vascular involvement and tumour resectability in pancreatic cancer, while RAD-FRS predicts clinically relevant postoperative pancreatic fistula risk from preoperative CT alone. Part III addresses the barriers to safe deployment through spatially-adaptive conformal prediction for uncertainty quantification, rule-based decision deferral to flag anatomically implausible outputs, and evaluation of model integration within hospital radiological workflow by expert radiologists.
Together, these studies show that bridging the persistent gap between AI research and clinical adoption depends on treating technical innovation, clinical value, and reliability as interdependent requirements. The thesis offers both methodological contributions and practical insight to guide the responsible integration of AI into hepato-pancreatic imaging.