Abstract 5424: A deep learning approach (AI) which accurately identifies breast tumor cells, tumor infiltrating lymphocytes (TILS) and fibroblasts from H&E slides

Abstract Background: Histopathology assessments of cancers require highly skilled pathologists, are labor intensive and prone to errors without proper training or fatigue. Machine learning can assist pathologists by increasing efficiency and minimizing individual variability. This study adapted a deep learning model to reliably identify 3 cell types from triple negative breast cancers fixed on H&E slides and verified performance with an expert pathologist. Methods: We apply the U-Net architecture to analyse pathology slides fixed with triple negative breast cancer (TNBC) tissue. The public dataset NuCLS was used for training. Semantic segmentation was used to identify single cells of 3 types: tumor cells, fibroblasts, and tumor infiltrating lymphocyte (TILs). For validation, the pathologist annotated 8 random H&E tiles. The 3 cell types accuracy for NuCLS and the model was evaluated by our pathologist. Results: Overall, there was a 73% agreement with the pathologists (Pathologist vs. NuCLS). A set of 1,555 (90%) TNBC slides were used for training and 173 for validation (10%; unseen data). Table 1 outlines the accuracy metrics for each cell type and for each comparison. Compared to our pathologist, the model accurately identified TILs (62%), followed by fibroblasts (42%) and lastly tumor cells (26%). A significant source of discrepancy was variation in labeled single cell boundaries. The model was better at identifying TILs. The pathologist took 1.5 hours to annotate 8 tiles for the 3 cells and our model 644ms. Conclusion: It is possible to develop a deep learning model to identify breast cancer cells, fibroblasts and TILs from H&E stained slides, with similar accuracy levels as a trained pathologist. The model performed better than a pathologist in identifying TILs, but both struggled with fibroblasts. Accuracy of 71% overall and 87% for TILs, motivates expansion to further datasets and other cancer types. Table 1. - Accuracy metrics. Quality assessment8 H&E tiles Pathologist vs. NuCLS U-Net model vs. NuCLs U-Net model vs. Pathologist U-Net model vs. NuCLs (Full) Background 86% 71% 62% 73% Tumour 43% 35% 26% 71% Fibroblast 45% 39% 42% 31% TILs 34% 96% 62% 87% Overall 73% 74% 61% 71% Citation Format: Luiz Augusto Zillmann da Silva, Alistair R. Williams, Aidan Kubeyev, Andrea Giorni, Jordan Laurie, Prabu Sivasubramaniam, Matthew Foster, Matthew Griffiths, Uzma Asghar. A deep learning approach (AI) which accurately identifies breast tumor cells, tumor infiltrating lymphocytes (TILS) and fibroblasts from H&E slides. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5424.