Cross-Spatial-Scale Radiomics for Predicting Pathologic Complete Response to Neoadjuvant Chemotherapy in Breast...

Pathologic complete response (pCR) is a key endpoint after neoadjuvant chemotherapy (NAC) in breast cancer, but whole-tumor radiomics may dilute heterogeneous response signals. We developed and externally validated TriSpace, an interpretable MRI biomarker integrating global tumor, intratumoral, peritumoral, and clinicopathologic information for pretreatment pCR prediction.

This retrospective four-center study included 658 patients who underwent pretreatment dynamic contrast-enhanced T1-weighted MRI before standard NAC (February 2018-April 2025). Patients from two centers formed training (n = 276) and internal validation (n = 117) sets; two other centers formed external test sets (n = 134 and n = 131). TriSpace fused whole-tumor radiomics, habitat-based intratumoral heterogeneity, the optimal peritumoral infiltration zone, and clinicopathologic features. Performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration, decision curve analysis, and SHAP.

Discrimination showed an inside-out hierarchy: intratumoral heterogeneity was the strongest single scale (external AUCs, 0.813 and 0.805), the peritumoral signal peaked at 3 mm and declined with distance, and whole-tumor radiomics was least robust (external AUCs, 0.628 and 0.597). TriSpace achieved the highest AUCs in all cohorts (0.946, 0.887, 0.852, and 0.815), with good calibration (all Hosmer-Lemeshow P >.05) and greater net benefit. SHAP indicated that imaging features compensated when clinicopathologic variables weakened under cross-center distribution shift; discrimination was maintained across nine subgroups (AUC range, 0.778-0.851).

TriSpace may improve interpretable pretreatment pCR prediction by integrating complementary tumor information across spatial scales, with consistent external performance.