Evidence Structures for Clinical Reasoning in Medical Imaging
A clinician does more than name what an image shows. They situate a finding in anatomy, weigh it against the patient's other findings, and hold back when the evidence is thin. This thesis argues that these acts of judgment are missing from medical imaging models because standard training rewards only the final prediction. It organizes them into a common framework, shows through three architectures spanning chest radiographs, CT, and dermatology that building these priors in improves how well models perform these operations, and dissects the representations current systems learn across radiology and histopathology to pinpoint where they fall short.
