Date of Award

12-2023

Document Type

Thesis

Degree Name

Master of Science in Operations Research

Department

Department of Operational Sciences

First Advisor

Nathan B. Gaw, PhD

Abstract

Saliency maps are a widely used methodology to make deep learning models more interpretable. They provide post-hoc explanations through identification of the most pertinent areas of an input medical image. These techniques have been assessed based on 1) localization utility, 2) sensitivity to model weight randomization, 3) repeatability, and 4) reproducibility. Ten saliency map techniques will be tested, Grad, Smooth Grad, Integrated Gradients, Smooth Integrated Gradients, XRAI, Grad-CAM, Guided Backpropagation, Guided GradCAM, GradCAM++, and ScoreCAM. The neural networks used to predict and read medical information require a reliable solution to provide medical practitioners intelligible results. Using the information of two, large, publicly available radiology datasets, we look to quantify and posit a solution to this problem so that the usage of saliency maps in the high-risk domain of medical imaging will require less scrutiny and can hold a more trustworthy base.

AFIT Designator

AFIT-ENS-MS-23-D-010

Comments

A 12-month embargo was observed for this thesis.

Approved for public release; PA case number on file.

Share

COinS