CMU-CS-22-137
Computer Science Department
School of Computer Science, Carnegie Mellon University



CMU-CS-22-137

Interpretability Approaches for a Breast Lesion Detection Model

Umaymah Imran

M.S. Thesis

August 2022

CMU-CS-22-137.pdf


Keywords: Breast lesion detection, interpretability, radiomic features, deep learning, feature analysis

Medical imaging holds an important role due to its ability to non-invasively visualize and analyze the internal structures of the human body. The rise in medical imaging data is putting an increased pressure on physicians/radiologists to efficiently perform clinical imaging tasks, which has in turn driven the need for development of diagnosis tools in healthcare. Therefore, over the last decade, there have been significant breakthroughs in the field of artificial intelligence for healthcare. Of these breakthroughs, the most important would be the development and deployment of deep neural networks (DNNs). These DNNs have a complex structure and consist of several computation layers. Their complex structure is what owes their ability to resolve challenging imaging tasks. However, several issues have been raised considering the black box nature of deep learning algorithms, which is why regardless of their impressive performance, DNNs have not achieved significant deployment in medical practice.

Specifically in healthcare, deep learning algorithms cannot be used for patient care unless the reasoning behind their outputs is explained due to the high stakes involved. Considering this, model interpretability is very important. It helps identify hidden information from the medical imaging data that may otherwise be invisible to the human eye. Model interpretability also adds to the trust of healthcare providers and patients involved. Beyond interpretability, analyzing model performance on targeted features could also allow model developers to debug their models.

In this thesis, we perform an analysis of a breast lesion detection model using different interpretability approaches. More specifically, we perform a preliminary analysis using the existing data and metadata. Then, we analyze hidden features for medical imaging data–that is, radiomic features–to further investigate model performance on certain input features. Finally, we determine the impact of the input data's location from the mammogram and suggest some future methods to improve our existing approaches

50 pages

Thesis Committee:
Adam Perer (Chair)
Kenneth Holstein

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science


Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu