Multiparametric MRI and transfer learning for predicting positive margins in breast-conserving surgery: a multi-center study

Abstract

This study aimed to predict positive surgical margins in breast-conserving surgery (BCS) using multiparametric MRI (mpMRI) and radiomics. A retrospective analysis was conducted on data from 444 BCS patients from three Chinese hospitals between 2019 and 2024, divided into four cohorts and five datasets. Radiomics features from preoperative mpMRI, along with clinicopathological data, were extracted and selected using statistical methods and LASSO logistic regression. Eight machine learning classifiers, integrated with a transfer learning (TL) method, were applied to enhance model generalization. The model achieved an AUC of 0.889 in the internal test set and 0.771 in the validation set. Notably, TL significantly improved performance in two external validation sets, increasing the AUC from 0.533 to 0.902 in XAH and from 0.359 to 0.855 in YNCH. These findings highlight the potential of combining mpMRI and TL to provide accurate predictions for positive surgical margins in BCS, with promising implications for broader clinical application across multiple hospitals.

Keywords: breast-conserving surgery; multiparameter magnetic resonance imaging; surgical margins; transfer learning.