Context Clues: Identifying Solutions to an Explainability Barrier for the Clinical Implementation of Machine Learning Models for the Case of Prostate Cancer

Despite many advances in biomedical engineering and clinical research, methods for cancer diagnoses have remained constant and invasive. Machine Learning has grown in popularity due to the potential to offer a noninvasive method to both diagnose cancer from imaging and predict cancer severity through identification of imaging data. Many previous published algorithms for cancer severity, however, have not been replicable or explainable and provide no context for their selections. Explainability is important as a model for prostate cancer must ensure that contextual information, such as the location of the prostate tumor, is accurately reflected in model selections. To fix these issues, this work a (1) evaluated the current state of model explainability within machine learning for cancer imaging through a literature review and creation of a series of predictive models for clinically significant prostate cancer, (2) developed a custom quantitative metric to evaluate the explainability of the features used in a deep learning model, and (3) evaluated this custom quantitative metric on prostate cancer medical imaging data. This explainable methodology is unique in that it provides an in-model method to easily calculate if the most clinically important region of anatomic abnormality co-localizes with the most-important regions as determined by the predictive model, whereas other explainable methods for machine learning models are often qualitative and require post-hoc analyses. The custom explainability metric developed using the most-important feature map of the image classification model provides a unique method to quantitatively analyze the quality of included features in a deep learning model.

Author

• Provenzano, Destie Jo

Language

• English

Keyword

• Medical Imaging
• Machine Learning
• Cervix Cancer
• Residual Neural Network
• Prostate Cancer
• Convolutional Neural Network

Date created

• 2025

Type of Work

• Dissertation

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Biomedical Engineering

Degree

• Ph.D.

Advisor

• Loew, Murray

Committee Member(s)

• Pless, Robert
• Hyuk Park, Chung
• Zara, Jason
• Haji-Momenian, Shawn

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/wd375x259