Differentiation of Benign and Malignant Breast Lesions Based on Multimodal MRI and Deep Learning

Abstract

To improve the accuracy of computer aided diagnosis (CAD) based on dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) in the differentiation of benign and malignant breast lesions, this study proposed a convolutional neural network model (AC_Ulsam_CNN) that is based on multi-modal feature fusion and the combination of asymmetric convolution (AC) and ultra-lightweight subspace attention module (Ulsam). Firstly, the transfer learning method was used to pre-train the model to screen out the most effective DCE-MRI time phase scan for differentiating benign and malignant breast lesions. Then, a network model based on AC_Ulsam_CNN was constructed based on the optimal time phase scan images to enhance the feature expression ability and robustness of the classification model. Finally, multimodal information such as breast imaging reporting and data system (BI-RADS) classification, apparent diffusion coefficient (ADC) and time-signal intensity curve (TIC) type were incorporated for feature fusion, to further improve the distinguishing performance of benign and malignant breast lesions. The performance of the model was verified by 5-fold cross-validation method, and the accuracy (ACC) of the proposed method was 0.826 and the area under the curve (AUC) was 0.877. The experimental results show that the proposed algorithm performs well in the classification of benign and malignant breast lesions with small sample size, and the fusion model based on multimodal data further enriches the feature information, thus this study improves the detection accuracy of lesions, and provides a new method for automatic differential diagnosis of benign and malignant breast lesions.

Keywords

• dynamic contrast enhanced magnetic resonance imaging
• convolutional neural network
• multimodal feature fusion
• breast lesions
• differentiation of benign and malignant lesions