Patch work.

With PSA screening coming back in vogue, pathologists everywhere may soon find themselves Gleason-ed out. This high-tech British Columbian comparative analysis examines the best way we should grade artificial intelligence (AI) on its ability to grade prostate cancer. Six “expert” pathologists independently assigned single Gleason scores 3-5 or benign to 333 snippets (aka “patches”) at 40x magnification from digitized histopathology slides from 231 prostatectomy samples. The majority vote on each snippet was used to develop CNN-trained AI. Anyone who actually understands the methods section probably dropped out of college and moved to Silicon Valley long before med school, but we’ll do our best explaining the results. Accuracies of AI versus pathologist at the snippet level were fantastic, but they started to fall short at the whole-core and especially patient levels. In other words, AI was prone to missing the forest for the trees. TBL: This is a cautionary tale warning reported accuracies of AI-generated path reports are only clinically-meaningful at the patient (i.e., not snippet) level. | Nir, JAMA Netw Open 2019