MR imaging of breast cancer: Interpretable radiomics analysis to assess treatment response and survival prognosis after neoadjuvant therapy

Abstract The early prediction of pathological complete response (pCR) status after neoadjuvant chemotherapy (NAC) determines individualized treatment and prognosis in breast cancer patients. The study aimed to develop a biologically interpretable MRI‐based radiomics model for predicting pCR. A total of 843 eligible breast cancer patients were retrospectively recruited from three medical centers and further divided into training, internal, and external validation cohorts. Radiomics features were extracted from tumors and 4 mm peritumor regions on dynamic contrast‐enhanced (DCE)‐MRI images. Independent clinical‐pathological factors and important radiomics features were determined step by step; then a pCR‐related hybrid model was constructed with AUCs of 0.802 to 0.893, which significantly outperformed the junior (AUCs of 0.623 to 0.761) and senior radiologists (AUCs of 0.648 to 0.771) with both DeLong test p < 0.05 in three cohorts. The prognostic stratification efficacy associated with disease‐free survival was demonstrated in three cohorts with both log‐rank p < 0.05. In terms of interpretability, the prediction contribution of the important radiomics features was first quantified by SHapley Additive exPlanation analysis. Furthermore, 120 eligible patients with DCE‐MRI and bulk‐RNA information were recruited from the TCGA‐BRCA cohort to explore the underlying biological information for the radiomics model. Radiogenomics analysis revealed that the predicted pCR group has active immune and inflammatory pathways, while the non‐pCR group performed a stronger cell proliferation and tumor invasion processes. In conclusion, these findings established an MRI‐based biologically interpretable pCR predictive tool, which has the potential to individually guide clinical decision making.