Explainable AI-Derived Spatial Pathological Features of Tumor, Necrosis, and Lymphocytes Identify Key Histological Signatures for Residual Cancer Burden Assessment in Breast Cancer

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background: Conventional histopathological assessment of residual cancer burden (RCB) following neoadjuvant therapy for breast cancer relies on manual visual evaluation, lacking objective quantification of spatial interactions among tumor, necrosis, and lymphocyte compartments. This limits prognostic accuracy and impedes personalized treatment stratification. Methods: We developed an explainable AI framework to segment three tissue compartments from H&E patches and constructed 11 spatial pathological features. This two-center study included 347 post-neoadjuvant patients (182 development, 165 external validation) plus 843 TCGA-BRCA cases for exploratory external validation. VGG-16, ResNet-50, and ViT-Base were evaluated; multiple instance learning identified target regions for feature construction. Six machine learning models with LASSO-selected features were interpreted via SHAP and Grad-CAM, with TCGA-BRCA serving as an exploratory external cohort. Results: VGG-16 achieved optimal segmentation (all AUCs > 0.90). In the external validation set, XGBoost achieved AUCs of 0.781 for Task 1 (RCB-0 vs. RCB-1+2+3) and 0.896 for Task 2 (RCB-0+1 vs. RCB-2+3). For Task 3 (RCB-0+1+2 vs. RCB-3), the model achieved an AUC of 1.000, with complete separation between groups. SHAP analysis revealed that the Necrotic Tumor Mixed Cluster (T+N) was the dominant indicator for Task 1 (mean |SHAP| = 0.439) and Task 2 (mean |SHAP| = 0.803), while the Pure Necrotic-Rich Zone (N-only) was the dominant indicator for Task 3 (mean |SHAP| = 1.967). In exploratory TCGA association analyses, the Pure Tumor-Rich Zone (T-only) exhibited a positive correlation (ρ = 0.706, p < 0.001; T-only is a feature used to derive this model-estimated RCB-like score), whereas the Lymphocyte-Infiltrated Necrotic Zone (N+L) showed a negative correlation (ρ = −0.407, p < 0.001) with the model-estimated RCB-like score, linking tumor-rich zones to immunosuppression and lymphocyte-infiltrated necrotic zones to immune activation. Conclusions: Our spatial pathological features effectively estimate RCB, with T+N and N-only as the predominant SHAP-identified indicators across the three tasks. Exploratory analyses further suggest that tumor-rich and lymphocyte-infiltrated necrotic zones correlate with opposing immune phenotypes, supporting personalized breast cancer management.