Deep learning (DL) based applications have successfully solved numerous problems in machine perception. In radiology, DL-based image analysis systems are rapidly evolving and show progress in guiding treatment decisions, diagnosing, localizing disease on medical images, and improving radiologists' workflow. However, many DL-based radiological systems fail to generalize when deployed in new hospital settings, and the causes of these failures are not always clear. Although significant effort continues to be invested in applying DL algorithms to radiological data, many open questions and issues that arise from incomplete datasets remain. To bridge the gap, we first review the current state of artificial intelligence applied to radiology data, followed by juxtaposing the use of classical computer vision features (i.e., hand-crafted features) with the recent advances caused by deep learning. However, using DL is not an excuse for a lack of rigorous study design, which we demonstrate by proposing sanity tests that determine when a DL system is right for the wrong reasons. Having established the appropriate way to assess DL systems, we then turn to improve their efficacy and generalizability by leveraging prior information about human physiology and data derived from dual energy computed tomography scans. In this dissertation, we address the gaps in the radiology literature by introducing new tools, testing strategies, and methods to mitigate the influence of dataset biases.

Library of Congress Subject Headings

Deep learning (Machine learning)--Evaluation; Medical radiology; Data sets

Publication Date


Document Type


Student Type


Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)


Christopher Kanan

Advisor/Committee Member

Qi Yu

Advisor/Committee Member

Nathan Cahill


RIT – Main Campus

Plan Codes