TLScope: a deep learning framework for quantifying tertiary lymphoid structures from H&E images reveals prognostic heterogeneity across breast cancer subtypes.

Tertiary lymphoid structures (TLSs) are ectopic immune aggregates associated with antitumor immunity and favorable prognosis in various cancers. However, standardized approaches for TLS detection and quantification in breast cancer remain underdeveloped. We present TLScope, a deep learning-based framework that accurately identifies and quantifies TLSs in hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) of breast cancer. TLScope combines cell-level classification, tumor and adipose region segmentation, and biologically informed TLS validation to enable scalable analysis of TLS density. Applied to over 1000 WSIs from internal and TCGA cohorts, TLScope revealed significant associations between TLS density and clinicopathological features. TLSs were more frequently observed in tumors with higher histological grades and elevated Ki-67 expression. Moreover, patients with more TLSs exhibited improved overall survival. Although TLSs were most abundant in HER2-enriched and basal-like subtypes, their prognostic value was most pronounced in Luminal B tumors, suggesting a context-dependent role of TLSs in breast cancer. This work provides a standardized and interpretable tool for TLS assessment in breast cancer, facilitating deeper insights into tumor-immune interactions and patient outcomes.