Author Type

Graduate Student

Date of Award

Spring 4-22-2026

Document Type

Thesis

Publication Status

Version of Record

Submission Date

May 2026

Department

Physics

College Granting Degree

Charles E. Schmidt College of Science

Department Granting Degree

Physics

Degree Name

Professional Science Masters (PSM)

Thesis/Dissertation Advisor [Chair]

Andreas Kyriacou

Abstract

Breast cancer remains one of the leading causes of cancer-related death among women. Early detection through screening mammography is therefore very important. However, mammogram interpretation can be difficult because abnormalities may be subtle and spread across different views. The increasing workload of radiologists can also affect timely analysis. Artificial intelligence (AI) has been widely explored to support mammography interpretation by improving speed and consistency. However, many existing AI methods focus on patch-level or image-level analysis rather than the full patient exam. This limits their ability to capture the complete clinical picture, since important information is distributed across multiple views. In addition, many methods do not provide clear laterality information, which reduces their clinical value.

This thesis presents a patient-level framework for breast cancer classification from multi-view mammograms. The model learns from the full exam rather than isolated patches or single images by using multiple-instance learning. It classifies each patient exam as normal or abnormal and, for abnormal exams, also identifies the abnormal side. The dataset included 1,775 mammography studies, with 1479 normal and 296 abnormal cases. Experimental results showed that Linear Discriminant Analysis (LDA) provided the clearest separation between normal and abnormal studies. Gradient Boosting achieved the highest test accuracy of 88.4% with an Area Under the Curve (AUC) of 69.1%. Random Forest (RF) achieved a similar test accuracy of 88.2% and the highest test AUC of 70.2%. Overall, the proposed framework supports patient-level classification while providing clinically relevant laterality information.

Included in

Physics Commons

Share

COinS