Teamwork makes the dream work.

Top Line: Pathologists are getting a good run for their money when it comes to accurately classifying prostate cancers. 

The Study: This study used prostate biopsies from 1243 patients at a single Dutch institution. Data was extracted from the clinical pathology reports (no one specific "expert" reviewer—a big difference from the prior study). A single slide from each patient with the most aggressive or most prevalent disease was then used for training. Training was separated into 4 steps (much like the prior study): segmentation of tissue, segmentation of tumor, segmentation of normal epithelial tissue, and finally assignment of Gleason score. Once again, tumors with the same primary and secondary Gleason score (e.g. 3+3) were used to train the deep neural network. This model also had a 1% error rate for benign versus malignant. While this model performed slightly worse on its internal test set than the previous expert model, it ended up largely outperforming a group of expert pathologists at assigning Gleason grade in an external validation test with a mean kappa statistic of 0.85. Perhaps this model that learned from an entire institution’s pathology reports had more robust training setting it up to perform better on external data.

TBL: Another study demonstrates expert-level performance of a deep-learning model to diagnose and grade prostate cancer. | Bulten, Lancet Oncol 2020