Bridging the gap between prostate radiology and pathology through\n machine learning

Abstract

Prostate cancer is the second deadliest cancer for American men. While\nMagnetic Resonance Imaging (MRI) is increasingly used to guide targeted\nbiopsies for prostate cancer diagnosis, its utility remains limited due to high\nrates of false positives and false negatives as well as low inter-reader\nagreements. Machine learning methods to detect and localize cancer on prostate\nMRI can help standardize radiologist interpretations. However, existing machine\nlearning methods vary not only in model architecture, but also in the ground\ntruth labeling strategies used for model training. In this study, we compare\ndifferent labeling strategies, namely, pathology-confirmed radiologist labels,\npathologist labels on whole-mount histopathology images, and lesion-level and\npixel-level digital pathologist labels (previously validated deep learning\nalgorithm on histopathology images to predict pixel-level Gleason patterns) on\nwhole-mount histopathology images. We analyse the effects these labels have on\nthe performance of the trained machine learning models. Our experiments show\nthat (1) radiologist labels and models trained with them can miss cancers, or\nunderestimate cancer extent, (2) digital pathologist labels and models trained\nwith them have high concordance with pathologist labels, and (3) models trained\nwith digital pathologist labels achieve the best performance in prostate cancer\ndetection in two different cohorts with different disease distributions,\nirrespective of the model architecture used. Digital pathologist labels can\nreduce challenges associated with human annotations, including labor, time,\ninter- and intra-reader variability, and can help bridge the gap between\nprostate radiology and pathology by enabling the training of reliable machine\nlearning models to detect and localize prostate cancer on MRI.\n