(UroToday.com) Histopathologic grading for Non-muscle-invasive bladder cancer (NMIBC) guides the management of these disease, yet subjective criteria and poor reliability limit its value. In this study, the investigators developed a reproducible, quantitative, and explainable grading algorithm. This algorithm was then correlated with clinical outcomes.
Histologic image analysis was performed on 641 tissue microarray (TMA) cores from transurethral resections. Features of nuclear size, shape, alignment, and mitotic rate were extracted using Visiopharm software and analyzed individually. These features were then combined into regression and random forest models to establish thresholds between grades. Cox proportional hazards analysis was performed to identify prognostic histological features. Whole slide grade and stage were significantly associated with RFS. Uni- and multivariate models using histologic features differentiated low- and high-grade NMIBC. Variation in nuclear area alone distinguished between grades with 82% accuracy. Complex models combining features increased accuracy to 88%.
To adapt these findings to whole slide images, the investigators used deep learning to identify hotspots that are enriched for high-grade features. By identifying mitotic index and variation in nuclear size as key metrics, this work should allow pathologists to consider grade as a continuous variable that can be tuned to important outcomes like recurrence and progression.
Presented by: David Berman, MD, PhD, Queens University
Written by: Roger Li, MD, Urologic Oncologist, Moffitt Cancer Center, during the International Bladder Cancer Network Annual Meeting, September 28-October 1, 2022, Barcelona, Spain