AI-based framework pipeline for stromal tumor infiltrating lymphocytes quantification in H&E stained breast cancer histopathological images

Identifier:  imarina:9525038

Handle: https://hdl.handle.net/20.500.11797/imarina9525038

Authors:  Abdelazeez, Shrief; Fiorin, Alessio; Della Mea, Vincenzo; Adalid, Laia; Siemion, Krzysztof; Lopez, Carlos; Lejeune, Marylene; Korzynska, Anna

Abstract:
The tumor microenvironment (TME) plays a pivotal role in immuno-oncology, influencing cancer progression and treatment outcomes. Among its components, stromal tumor-infiltrating lymphocytes (sTILs) represent important prognostic and predictive biomarkers in breast cancer. However, visual assessment of sTILs is complex, time-consuming, and subject to inter-observer variability. To address these challenges, we propose an AI-based framework for sTILs quantification in hematoxylin and eosin (H&E)-stained breast cancer histopathological images across different molecular subtypes, in accordance with the recommendations of the International Immuno-Oncology Biomarker Working Group (IIOBWG). The framework integrates two parallel and explainable segmentation models trained on multi-center data: an attention-based model for TILs detection and a stroma segmentation model designed to identify both tumoral and infiltrated stromal regions as a unified clinically relevant compartment. The resulting segmentation outputs are combined through a scanner resolution-aware quantification algorithm, enabling robust sTILs scoring across data acquired from different scanners and institutions. The proposed framework was externally validated on two independent private datasets comprising 60 and 112 whole-slide images with expert-annotated Region Of Interests (ROIs). Strong agreement with expert pathologist scoring was achieved (Pearson correlation up to 0.73; Concordance Correlation Coefficient (CCC) up to 0.72), and the framework consistently outperformed state-of-the-art approaches when evaluated on the same data. These results demonstrate that the proposed framework provides a reliable, efficient, and scalable solution for standardized sTILs quantification, supporting its potential utility in both clinical practice and translational research.