Journal of Shanghai Jiao Tong University ›› 2023, Vol. 57 ›› Issue (5): 570-581.doi: 10.16183/j.cnki.jsjtu.2022.088

Special Issue: 《上海交通大学学报》2023年“生物医学工程”专题

• Biomedical Engineering • Previous Articles     Next Articles

UConvTrans:A Dual-Flow Cardiac Image Segmentation Network by Global and Local Information Integration

LI Qing1,2, HUANGFU Yubin1, LI Jiangyun1,2(), YANG Zhifang1, CHEN Peng3, WANG Zihan1   

  1. 1. School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, China
    2. Key Laboratory of Knowledge Automation for Industrial Processes of the Ministry of Education, University of Science and Technology Beijing, Beijing 100083, China
    3. FINTECH Innovation Division, Postal Savings Bank of China, Beijing 100808, China
  • Received:2022-03-31 Revised:2022-05-19 Accepted:2022-05-24 Online:2023-05-28 Published:2023-06-02
  • Contact: LI Jiangyun E-mail:leejy@ustb.edu.cn.

Abstract:

Cardiac magnetic resonance image (MRI) segmentation has the features such as there is a lot of noise, the target areas are indistinguishable from the background, and the shape of the right ventricle is irregular. Although convolution operations are good at extracting local features, the U-shaped convolutional neural networks (CNN) structure hardly models long-distance dependency between pixels and can not achieve ideal segmentation results on cardiac MRI. To solve these problems, UConvTrans is proposed with a dual-flow U-shaped network by global and local information integration. First, the network applies the CNN branch to extract local features and capture global representations by Transformer branch, which retains local detailed features and suppresses the interference of noise and background features in cardiac MRI. Next, the bidirectional fusion module is proposed to fuse the features extracted by CNN and the Transformer with each other, enhancing the feature expression capability and improving the segmentation accuracy of the right ventricle. Besides, the parameters of network can be set flexibly because the transformer structure in the proposed method does not require pre-trained weights. The proposed method also strikes a better balance between precision and efficiency, which is evaluated on the MICCAI 2017 ACDC dataset. The results show that the network outperforms U-Net by 1.13% average dice coefficient while the parameter amount and the floating point operations are only 10% and 8% of the U-Net. Finally, the proposed method achieves a dice coefficient of 92.42% for the right ventricle, 91.64% for the myocardium, and 95.06% for the left ventricle respectively and wins the first place in the myocardium and left ventricle on test set.

Key words: medical image segmentation, cardiac magnetic resonance image (MRI), convolutional neural network (CNN), Transformer models, encoder-decoder

CLC Number: