AI- Enabled Segmentation of Pathology Images for Prognostic Implications in Triple-Negative Breast Cancer

Abstract

Tumor-infiltrating lymphocytes (TILs) are critical prognostic indicators in triple-negative breast cancer (TNBC), with higher TIL levels correlating with better neoadjuvant chemotherapy (NAC) response and survival outcomes. Stromal TILs (sTILs) have shown particularly strong prognostic value, but consistent and reproducible sTIL scoring remains challenging due to subjectivity of manual evaluation. To address this, an AI-enabled pipeline called TIL segmentation (TILseg) was developed to generate sTIL scores from hematoxylin-and-eosin-stained (H&E-stained) whole-slide images (WSIs). TILseg processes WSIs by dividing them into patches, classifying tissue regions using a VGG19-based 3-class classifier (3CC), and segmenting lymphocyte nuclei with a pre-trained StarDist model. sTIL scores are computed as the proportion of lymphocyte area within the stromal mask, following the 2014 International TILs Working Group guidelines. Spatial heterogeneity was further analyzed by computing sTILs relative to the nearest tumor boundaries, yielding a spatial sTIL score. Our computational scores showed moderate correlation with pathologist-assessed sTIL scores but outperformed in prognostic analyses. Non-spatial and spatial scores both successfully stratified overall survival and recurrence-free survival, and notably spatial scores significantly improved the stratification of overall survival. In conclusion, TILseg provides a reproducible and scalable tool for sTIL quantification and spatial analysis, enhancing post-NAC prognostic assessment in TNBC.