An independent assessment of an artificial intelligence system for prostate cancer detection shows strong diagnostic accuracy

• DOI: 10.1038/s41379-021-00794-x
• Publication Date: 2021-03-29T07:02:36Z
• Authors: Sudhir Perincheri; Angelique Wolf Levi; Romulo Celli; Peter Gershkovich; David Rimm; Jon Stanley Morrow; Brandon Rothrock; Patricia Raciti; David Klimstra; John Sinard
• Keywords: Prostate biopsy; Digital Pathology; Surgical pathology

Prostate cancer is a leading cause of morbidity and mortality for adult males in the US. The diagnosis prostate carcinoma usually made on core needle biopsies obtained through transrectal approach. These may account significant portion pathologists' workload, yet variability experience expertise, as well fatigue pathologist adversely affect reliability detection. Machine-learning algorithms are increasingly being developed tools to aid improve diagnostic accuracy anatomic pathology. Paige AI-based digital one such tool trained slide archive New York's Memorial Sloan Kettering Cancer Center (MSKCC) that categorizes biopsy whole-slide image either "Suspicious" or "Not Suspicious" prostatic adenocarcinoma. To evaluate performance this program secured, processed, independently diagnosed at an unrelated institution, we used review 1876 images (WSIs) from our practice Yale Medicine. categorizations were compared pathology originally rendered glass slides each biopsy. Discrepancies between categorization by manually reviewed pathologists with specialized genitourinary expertise. showed sensitivity 97.7% positive predictive value 97.9%, specificity 99.3% negative 99.2% identifying data set derived independent institution. Areas improvement identified Prostate's handling poor quality scans. Overall, these results demonstrate feasibility porting machine-learning algorithm institution remote its training set, highlight potential powerful workflow evaluation surgical practices.