Artificial intelligence–based detection of treatment-emergent metastatic neuroendocrine carcinoma of prostate.

36 Background: Metastatic prostate carcinoma is a highly heterogeneous disease with a wide morphological spectrum and disease outcomes. While some retain adenocarcinoma morphology, some undergo lineage plasticity and transform into treatment emergent neuroendocrine carcinoma (t-NEPC) with morphology similarity to small cell carcinoma and highly aggressive disease behavior. We developed a morphology-based artificial intelligence (AI) tool applicable in real-world metastatic biopsies to assist the diagnosis of t-NEPC on digital whole-slide images (WSIs). Methods: Using a previously presented deep-learning based supervised learning algorithm, which was trained with localized prostate specimens with NEPC morphology in whole slide images (WSIs) using a custom-designed UNet model, here we fine tuned this algorithm using 62 metastatic biopsies from our institution with WSIs of H&E sections (enriched for t-NEPC, n=15, 24%), bench marking with GU pathologist’s diagnosis (adenocarcinoma vs. poorly differentiated carcinoma vs. t-NEPC) to generate a continuous t-NEPC score (range 0-1, with 1 most likely represent t-NEPC) for each slide. The algorithm was further tested in a non-overlapping external cohort (West Coast Dream Team and PROMOTE cohorts) of 167 metastatic biopsies with H&E WSIs (including 7 t-NEPC, 4%). Results: In our institutional cohort for fine tuning (n=62), t-NEPC score was significantly higher in 15 cases with a pathology diagnosis of t-NEPC compared to those with adenocarcinoma or poorly differentiated carcinoma. Using t-NEPC score >0.05 as a cutoff, the concordance rate between the AI t-NEPC score and pathologists was 94% t-NEPC and 56% for adenocarcinoma. In the validation cohort (West Coast Dream Team and PROMOTE, n=167), the algorithm still separated t-NEPC from adenocarcinoma, but not significantly in poorly differentiated carcinoma, with a concordance rate 71% for t-NEPC 77% for adenocarcinoma. Conclusions: As a proof-of-principle study, a morphology-based AI t-NEPC score established from a localized NEPC classifier was applicable towards the diagnosis of t-NEPC with a high concordance rate with GU pathologists. Adenocarcinoma Poorly differentiated Carcinoma t-NEPC Institutional cohort (n=62) N, % 34, 55% 12, 19% 15, 24% t-NEPC score (mean, +/- STD) 0.14, 0.20 0.09, 0.15 0.38, 0.21 Concordance (n, %) 19/34, 56% N/A 14/15, 94% West Coast Dream Team and PROMOTE cohorts (n=162) N, % 100, 62% 60, 37% 7, 4% t-NEPC score (mean, +/- STD) 0.05, 0.12 0.08, 0.17 0.24, 0.33 Concordance (n, %) 77/100, 77% N/A 5/7, 71%