Abstract

Info

Title:: Deep learning-based classification for benign and malignant breast masses using multimodal ultrasound images

Author(s):: WANG Yiwei1; LI Xiaobing1; NIE Shengdong1; JIANG Lixin2; WAN Caifeng2; JIANG Zhuoyun1; JIA Shouqiang3; 1. School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China 2. Department of Ultrasound, Renji Hospital, School of Medicine, Shanghai Jiaotong University, Shanghai 200025, China 3. Department of Imaging, Jinan Peoples Hospital Affiliated to Shandong First Medical University, Jinan 271100, China

Keywords:: Keywords: breast cancer deep learning ultrasound multimodality attention mechanism image classification

Abstract:: Abstract: A breast mass classification model based on dual neural networks and multimodal ultrasound images is proposed to address the issue of deficiency of information in monomodal ultrasound image. The 807 gray-scale images and 807 elastograms of 96 breast cancer patients (51 malignant and 45 benign) from Renji Hospital, School of Medicine, Shanghai Jiaotong University in 2021 were collected for experiment. The traditional ResNeXt101 model is improved by removing the final average pooling layer and fully connected layer, and adding an attention mechanism module to enhance the attention to important image information. Then, gray-scale images and elastograms of the same lesion of the patient are input into two improved ResNeXt101 networks separately and the features output from the two networks are concatenated and fused to construct a fully connected classification layer for the discrimination of benign and malignant breast masses. The experimental results show that the accuracy rate and AUC of dual neural network for multimodal ultrasound images are 84.27% and 0.932, higher than those of single neural network for monomodal ultrasound image.