Abstract
BACKGROUND: Peritumoral characteristics demonstrate significant predictive value for neoadjuvant chemotherapy (NAC) response in breast cancer (BC) through tumor-stromal interactions. Radiomics analysis of peritumoral regions has shown robust capability in predicting treatment outcomes; however, the optimal peritumoral thickness for maximizing predictive accuracy remains undefined. OBJECTIVE: To establish a clinically implementable framework for early identification of NAC non-responders through standardized prediction modeling. This study aims to determine the optimal peritumoral thickness for NAC response prediction by training and systematically comparing artificial intelligence (AI)-driven radiomics models across multiple peritumoral zones using Automated Breast Volume Scanning (ABVS). METHODS: A total of 402 BC patients who received NAC were retrospectively analyzed. Pre-treatment ABVS images were processed to extract radiomic features from five regions of interest (ROIs): the intratumoral region (R0) and four consecutive peritumoral zones (R2-R8) extending outward at 2-mm intervals. The study cohort was divided into training and testing cohorts. ROI-specific TabNet models were developed using the training cohort data. Comparative analysis was performed in the testing cohort through comprehensive performance evaluation, including discrimination, calibration, clinical utility assessment, and classification metrics, to identify the optimal peritumoral zone. The radiomics features of the best-performing model were ranked by importance, with subsequent ablation studies validating the predictive contribution of high-ranking features. RESULTS: Among the study population, 138 patients (34.3%) were classified as NAC non-responders. Model evaluation demonstrated progressively improved predictive performance from R0 to R6, with area under the ROC curves increasing from 0.681 to 0.845. The R6 model demonstrated optimal performance with accuracy of 0.810 and precision of 0.765. The combined model integrating R0 and R6 features enhanced predictive capability, achieving accuracy of 0.909, precision of 0.841, and recall of 0.902. Feature importance analysis identified textural heterogeneity and volumetric characteristics as the most influential variables, with the top features derived predominantly from the 6-mm peritumoral region. CONCLUSION: The 6-mm peritumoral zone demonstrated optimal predictive value for NAC response, with the AI-driven combined intratumoral-peritumoral model achieving superior performance. This standardized ABVS-based radiomics approach enables early identification of potential NAC non-responders, facilitating timely therapeutic modifications.