Author Type

Graduate Student

Date of Award

Summer 8-4-2026

Document Type

Thesis

Publication Status

Version of Record

Submission Date

August 2026

Department

Computer and Electrical Engineering and Computer Science

College Granting Degree

College of Engineering and Computer Science

Department Granting Degree

Electrical Engineering and Computer Science

Degree Name

Master of Science (MS)

Thesis/Dissertation Advisor [Chair]

Ionut Cardei

Abstract

Weakly supervised medical imaging models trained with coarse image-level labels often report strong performance on metrics such as accuracy and AUC. This thesis argues that these metrics can be misleading: a model may succeed on a coarse diagnostic task while failing to preserve the fine-grained information needed for consequential clinical decisions. It makes this failure measurable through the diagnostic gap, defined as the divergence between coarse and fine-grained diagnostic preservation, across three connected studies. The first shows that near-perfect ovarian ultrasound accuracy reflects visual separability rather than pathological understanding. The second measures the diagnostic gap in a mammographic pipeline, where coarse lesion-type features are preserved while malignancy information is lost. The third explains the mechanism: only about 4.4% of latent magnitude aligns with supervisory gradients. Before clinical deployment, models must therefore be evaluated at the appropriate level of diagnostic granularity, not just through simple accuracy.

Share

COinS