AI 驱动的 TIL（肿瘤浸润淋巴细胞）空间分析预测三阴性乳腺癌化疗-免疫治疗应答

We developed an integrated AI pipeline for automated analysis of hematoxylin-eosin (HE)-stained breast cancer samples. The framework incorporates: 1) TLS detection and classification, 2) TIL quantification, and 3) spatial relationship mapping between lymphocytes and tumor cells. Biomarkers associated with Miller-Payne (MP) response grades were identified using multivariate statistics, enabling construction of a random forest prognostic model.

The TLS recognition model demonstrated substantial agreement with pathologists ( =0.73), while the TIL classifier achieved 0.92 accuracy. Analysis of 32 HE images from triple-negative breast cancer (TNBC) cases treated with neoadjuvant chemo-immunotherapy revealed significant associations: stromal lymphocyte density and percentage positively correlated with MP grades (p<0.05); average cell counts in 2-cell and 3-cell lymphocyte aggregates showed no significant correlations; shorter mean minimum distances between lymphocytes and tumor cells within 10-30 m radius range were inversely associated with MP grades (p<0.05). The spatial-feature-based prediction model achieved an AUC of 0.81 (95% CI: 0.76-0.86).

This study established an AI-driven HE analysis pipeline that precisely quantified spatial determinants of tumor-immune interactions. Lymphocyte spatial organization, particularly proximity to tumor cells and lymphocyte aggregation patterns, serves as a critical predictor of chemo-immunotherapy response beyond density metrics. The validated MP-grade prediction model demonstrates translational potential for clinical decision-making in TNBC management, although external validation in larger cohorts is warranted.