Interpretable Deep Learning Model for Prostate Cancer Detection

Grade Level at Time of Presentation

Sophomore

Major

Computer Engineering & Computer Science

Institution

University of Louisville

KY House District #

28

KY Senate District #

37

Department

Computer Engineering & Computer Science

Abstract

Prostate cancer is the second leading cause of cancer death in American men, behind only lung cancer. Detecting prostate cancer early and accurately are key factors in preventing these deaths. Progress has been made in creating deep learning systems that are able to detect prostate cancer with a high degree of accuracy. However, an indispensable problem with these systems is while the performance can be exceptionally accurate, the classification outputs are non-interpretable. This non-interpretable characteristic significantly inhibits these models from being implemented in medical settings. We address this problem of interpretability of deep learning systems in the domain of prostate cancer detection. We develop a deep convolutional neural network based on the VGG16 architecture for the classification of prostate cancer lesions using T2 weighted magnetic resonance images. Our model achieves high level performance with an AUC of 0.86, sensitivity of 0.88, and specificity of 0.88. We use saliency maps for interpretation by calculating how much each individual pixel contributes to the overall class scores. We show the clusters of pixels that contribute the most to the prediction thus showing the reasoning behind the classification. We then show the interpretation caliber to demonstrate the exactness of the interpretation. This work demonstrates the potential to use saliency maps to interpret classifications of deep learning prostate cancer detection systems.

This document is currently not available here.

Share

COinS
 

Interpretable Deep Learning Model for Prostate Cancer Detection

Prostate cancer is the second leading cause of cancer death in American men, behind only lung cancer. Detecting prostate cancer early and accurately are key factors in preventing these deaths. Progress has been made in creating deep learning systems that are able to detect prostate cancer with a high degree of accuracy. However, an indispensable problem with these systems is while the performance can be exceptionally accurate, the classification outputs are non-interpretable. This non-interpretable characteristic significantly inhibits these models from being implemented in medical settings. We address this problem of interpretability of deep learning systems in the domain of prostate cancer detection. We develop a deep convolutional neural network based on the VGG16 architecture for the classification of prostate cancer lesions using T2 weighted magnetic resonance images. Our model achieves high level performance with an AUC of 0.86, sensitivity of 0.88, and specificity of 0.88. We use saliency maps for interpretation by calculating how much each individual pixel contributes to the overall class scores. We show the clusters of pixels that contribute the most to the prediction thus showing the reasoning behind the classification. We then show the interpretation caliber to demonstrate the exactness of the interpretation. This work demonstrates the potential to use saliency maps to interpret classifications of deep learning prostate cancer detection systems.