Integrating intratumoral and peritumoral radiomics with deep transfer learning from multiparametric MRI for preoperative prediction of HER2 status in breast cancer: a multicenter study

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Purpose To develop and validate a combined model integrating intratumoral and peritumoral radiomics features, deep transfer learning features from multiparametric MRI (DCE-MRI, T2WI, and DWI), and clinical indicators, and to evaluate its diagnostic performance and clinical utility for preoperative prediction of HER2 expression status in breast cancer. Methods We collected data from 411 breast cancer patients from three centers retrospectively (training set: 212; internal validation set: 91; external test sets: 50 and 58). Multiparametric MRI (DCE/T2WI/DWI) was acquired. This study extracted manually constructed radiomics features and deep transfer-learning features based on ResNet50 from the tumor interior and peritumoral regions using multiparametric MRI. Through multiple feature-selection steps, such as intraclass correlation coefficient calculation, Spearman correlation testing, and LASSO logistic regression, a fusion feature set of deep learning and radiomics (DLR) was constructed. Finally, the DLR feature set was combined with independent clinical predictors to establish a combined prediction model. The model performance was evaluated using receiver operating characteristic curves, calibration curves, and decision curve analysis, and visual interpretability analysis was conducted using Grad-CAM and SHAP methods. Results The combined model had the best accuracy and prediction ability, with AUC values of 0.965 (95% CI: 0.939–0.990) and 0.904 (95% CI: 0.843–0.966) for the training and internal validation cohorts, respectively. In external test sets 1 and 2, it had AUCs of 0.844 (95% CI: 0.724–0.964) and 0.846 (95% CI: 0.743–0.949), respectively. Conclusions By integrating intratumoral/peritumoral features from multiparametric MRI, radiomics, and deep transfer learning, and combining these with clinical indicators, the developed model enables precise prediction of HER2 status in breast cancer. This provides a reliable assessment tool for precision diagnosis and treatment decisions.