Abstract
BACKGROUND: Current methods utilizing preoperative magnetic resonance imaging (MRI)-based radiomics for assessing lymphovascular invasion (LVI) in patients with early-stage breast cancer lack precision, limiting the options for surgical planning. PURPOSE: This study aimed to develop a sophisticated deep learning framework called "Prior Clinico-Radiological Features Informed Multi-Modal MR Images Convolutional Neural Network (PCMM-Net)" to improve the accuracy of LVI prediction in breast cancer. By incorporating multiparameter MRI and prior clinical knowledge, PCMM-Net should enhance the precision of LVI assessment. METHODS: A total of 341 patients with breast cancer were randomly divided into training and validation groups at a ratio of 7:3. Imaging features were extracted from T1-weighted, T2-weighted, and contrast-enhanced T1-weighted MRI sequences. Stepwise univariate and multivariate logistic regression were employed to establish a clinico-radiological model for LVI prediction. The radiomics model was built using redundancy and the least absolute shrinkage and selection operator. Then, two deep learning frameworks were developed: the Multi-Modal MR Images Convolutional Neural Network (MM-Net), which does not consider prior radiological features, and PCMM-Net, which incorporates multiparameter MRI and prior clinical knowledge. Receiver operating characteristic curves were used, and the corresponding areas under the curves (AUCs) were calculated for evaluation. RESULTS: PCMM-Net achieved the highest AUC of 0.843. The clinico-radiological features displayed the lowest AUC value of 0.743, followed by MM-Net with an AUC of 0.774, and radiomics with an AUC of 0.795. CONCLUSIONS: This study introduces PCMM-Net, an innovative deep learning framework that integrates prior clinico-radiological features for accurate LVI prediction in breast cancer. PCMM-Net demonstrates excellent diagnostic performance and facilitates the application of precision medicine.