Abstract 2422: Fine-grained inflammatory cell segmentation in histopathology with deep learning
Pierre‐Antoine Bannier, Benjamin Adjadj, Sebastien Mandela, Guillaume Horent, Thomas Mathieu, Ulysse Marteau, Valentin Gaury, Laura Dumont, Aurore Lyon, Reda Belbahri, Benoît Schmauch, Kathryn Schutte, Lucie Gillet, Katharina von Loga, Caroline Hoffmann · Cancer Research · 2025
Abstract Introduction: Deep learning can advance clinical research by improving patient subtyping and uncovering biological insights. However, many models have limited interpretability, reducing clinical utility. To address this, we developed a model that detects, segments, and classifies cells on hematoxylin and eosin (H&E) slides, including understudied types like neutrophils, eosinophils, and plasmocytes, enabling novel biomarkers related to the tumor microenvironment (TME). Methods: We collected a novel training set composed of 1,479 patches of size 112μm x 112μm from H&E-stained TCGA slides across five indications: bladder cancer (n=400), colon adenocarcinoma (n=242), lung adenocarcinoma (n=231), lung squamous cell carcinoma (n=220), and mesothelioma (n=386). We collected two external validation sets including patches from muscle invasive bladder cancer (MIBC) (n=79) and mesothelioma (n=48) cohorts. For both training and validation, three expert pathologists annotated cells into seven classes: cancer cells, lymphocytes, fibroblasts, neutrophils, eosinophils, plasmocytes, and others. We used a consensus rule to resolve discrepancies at the cell level between pathologists. In total, 151,404 nuclei were annotated in the training set and 9,565 in the validation set. We trained a model in 3-fold cross-validation on the training set. We independently evaluated its performance on our validation sets and on the Lizard dataset (colorectal cancer). For each cell type we report the panoptic quality (PQ), which combines detection accuracy and segmentation quality. Results: Conclusion: Our model demonstrates consistent performance across multiple validation cohorts, including understudied cell types for which it even outperforms models trained in-domain on Lizard. This reliability highlights its potential for broader clinical applications. By enabling the identification of immune cells, the model paves the way for designing novel interpretable TME-related biomarkers. Citation Format: Pierre-Antoine Bannier, Benjamin Adjadj, Sebastien Mandela, Guillaume Horent, Thomas Mathieu, Ulysse Marteau, Valentin Gaury, Laura Dumont, Aurore Lyon, Reda Belbahri, Benoit Schmauch, Kathryn Schutte, Lucie Gillet, Katharina Von Loga, Caroline Hoffmann. Fine-grained inflammatory cell segmentation in histopathology with deep learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2422.