A modular deep learning pipeline for stromal TILs scoring in breast cancer H&E slides

Shrief Abdelazeez, Faisal Ahmed, Laia Adalid Llansa, Krzysztof Siemion, Carlos López, Marylène Lejeune, H. Rashwan, A. Korzyńska

TL;DR: This work presents a modular deep learning pipeline that delivers fully automated and continuous stromal TILs scores in line with the International Immuno-Oncology Biomarker Working Group (IIOBWG) guidelines and is interpretable: each stage produces human-readable outputs.

BACKGROUND AND OBJECTIVE Tumor-infiltrating lymphocytes (TILs) are an important indicator of immune activity in breast cancer, yet scoring them consistently on H&E slides remains challenging in routine pathology. This work presents a modular deep learning pipeline that delivers fully automated and continuous stromal TILs (sTILs) scores in line with the International Immuno-Oncology Biomarker Working Group (IIOBWG) guidelines. METHODS The pipeline combines three components: a TIL segmentation model refined through pathologist-guided active learning, a robust stroma segmentation network based on an enhanced DeepLabV3+, and a lightweight regression module that learns how TILs distribute within stromal regions. A new adaptive aggregation strategy integrates patch-level predictions into a single, clinically meaningful score while accounting for heterogeneous infiltration. RESULTS The system was evaluated on two independent datasets (60 and 112 WSIs) with expert-annotated ROIs, achieving strong agreement with pathologists (Pearson of 0.814; ICC of 0.808). CONCLUSIONS Importantly, the pipeline is interpretable: each stage produces human-readable outputs (stroma masks, TIL-in-stroma maps), and SegGradCAM visualizations confirm that predictions rely on biologically relevant tissue regions. These findings demonstrate the pipeline's potential as a reliable and clinically adaptable tool for standardized, fully automated TILs quantification in breast cancer pathology. The source code and pretrained models are publicly available at https://github.com/Shrief-Abdelazeez/TILs-Scoring.