Deep Learning for the Prognostication of Prostate Cancer

Purpose and goal

Our goal was to use deep learning to improve the CADESS prostate cancer prognostication classifier. We improved he classifiers sensitivity by 24%. The resulting sensitivity, that is the ability to identify cancer, is now 93.6% with a specificity of 90%, which far exceeds any published result.

Expected results and effects

In the last ten years, Active Surveillance has emerged as a primary management strategy in men with favorable-risk prostate cancer. It aims to delay or avoid curative treatment by repeated PSA testing and biopsies, thereby giving men more quality-of-life years. It relies on accurate risk assessment, in particular accurate malignancy grading. By improving its classifier, CADESS will improve the accuracy of prostate cancer risk assessment and will support active surveillance in an substantial portion of men afflicted with prostate cancer.

Planned approach and implementation

Deep learning requires very large samples of training data annotated by experts. Prostate tissue is heterogeneous: a small area of high-grade tissue may be surrounded by benign tissue, and conversely, large malignant areas may contain many different grades as well as benign tissue. A training data set needs to capture all these variations; just indicating a large area of glands, some of which are cancerous, will confuse the training process. CADESS technology allowed us to automatically create an annotated training data set, which is the secret to the improvement of our classifier.