Automated tumor annotation on digitized prostate cancer histology

What is it about?

In a cohort of 47 patients, we found that both machine and deep learning models were able to accurately detect regions of cancer from noncancer in prostate cancer histology. We specifically found that our ensemble model was better able to classify these regions, however, performed poorly compared to our residual network at detecting regions of high- and low-grade cancer from noncancer, as well as unique Gleason patterns.

Featured Image

Why is it important?

Prostate cancer is the most diagnosed non-cutaneous cancer in men, though not all cases are high risk. Prostate cancer is graded using the Gleason grading system, which assigns scores to the two most predominate morphological patterns present. These have recently been stratified into five Grade Groups to predict patient prognosis. While the Gleason grading system is the current gold-standard, it is a subjective, quantitative metric that suffers from interrater variability. Therefore, in this study we developed two models to quantitatively assign Gleason scores to whole slide prostate cancer histology: an (1) ensemble machine learning model using quantitative histomorphometric features as input, and (2) a residual network with 101 layers. We specifically tested the hypothesis that a machine learning model using calculated histological features would discriminate prostate cancer from normal tissue; however, a deep learning model would better classify the unique Gleason patterns.