Establishing a Multimodal Model for Predicting Lymphovascular Invasion in Breast Cancer Using Deep Transfer Learning Based on Ultrasound and Dynamic Contrast-Enhanced Magnetic Resonance Imaging.
Authors
Abstract
Develop a multimodal model based on ultrasound (US) and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) data for predicting lymphovascular invasion (LVI) in breast cancer. A total of 326 patients with breast cancer were enrolled in this retrospective study. All participants underwent preoperative US and DCE-MRI. Data from Center 1 (n=167) were divided into a training set and an internal validation set, while data from Center 2 (n=159) were used as the external validation set for analysis. A fusion feature set was constructed from intratumoral regions (ITR) of US and DCE-MRI using deep transfer learning (DTL) features. Among three machine learning models-Support Vector Machine (SVM), k-Nearest Neighbors (KNN), and Multi-Layer Perceptron (MLP)-the MLP classifier trained with a feature subset selected via LASSO achieved optimal performance across multiple feature selection methods and was thus designated as the final model. The diagnostic performance and clinical utility of the multimodal model were validated and assessed through the area under the receiver operating characteristic (ROC) curve and decision curve analysis (DCA). Model interpretability was assessed using SHapley Additive exPlanations (SHAP). Results indicate that the multimodal model, utilizing DTL features from US and DCE-MRI within the tumor region, demonstrate outstanding performance in classifying LVI in breast cancer. The AUC values of the model were 0.987 in the training set, 0.967 in the internal validation set and 0.942 in the external validation set. Its performance was significantly better than the models using DTL features derived from single-region US (AUC: 0.966, 0.852, 0.789) or DCE-MRI (AUC: 0.936, 0.716, 0.746), indicating great value for clinical application. Multimodal models based on DTL features can accurately distinguish LVI status in breast cancer patients. This model demonstrates promising application prospects in distinguishing LVI in breast cancer, with its value lying in effectively enhancing diagnostic accuracy and robustness.