Automated segmentation of the gastrocnemius and soleus in shank ultrasound images through deep residual neural network. (March 2022)

• Record Type: Journal Article
• Authors: Du, Getao; Zhan, Yonghua; Zhang, Yue; Guo, Jianzhong; Chen, Xueli; Liang, Jimin; Zhao, Heng
• Is Part Of: Biomedical signal processing and control. Volume 73(2022)
• Publication Date: 2022-03
• Subjects: Muscle atrophy -- Segmentation -- Deep learning -- Deep residual neural network -- Dilated convolution; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103447
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We developed a novel deep learning network to segment the gastrocnemius and soleus in shank ultrasound (US) images. As the sublayer unit of each layer of a U-net, deep Resnet was used to meet the challenges of poor US image quality and low contrast. Dilated convolution was used to extract deeper feature information, and then reduce the interference of a large number of shadows. This method has achieved a mean DSC of the Gas and Sol muscles of 94.82% and 90.72%, which outperforms other state-of-the-art methods. Abstract: Segmentation of the shank gastrocnemius (Gas) and soleus (Sol) muscles in ultrasound (US) images allows to extract the muscle features, which are important for the early diagnosis of muscle atrophy. The automatic segmentation of the muscles is a challenging task, and deep learning (DL) provides a solution to this problem, which can effectively extract representative features from the muscle regions and background of the images. In this study, we propose ResTU-net, an automatic segmentation method based on improved U-net network, to segment the Gas and Sol muscles in shank US images. This network uses the deep residual neural network (Resnet) as the sublayer unit of each layer of a U-net, and can effectively combine the features of each layer with those of the next layer to meet the challenges of poor US image quality and low contrast. In addition, dilated convolution is used instead of the pooling layer in the network to prevent information loss during training. Experiments were performed on 3350 shank US images from 23 Sprague Dawley (SD) rats, among them, 2650 shank US images were used for network training and 700 for network validation. Compared with state-of-the-art networks, the experimental results show that the method can achieved the best segmentation capability results and a mean Dice similarity coefficient (DSC) of the Gas and Sol muscles of 94.82% and 90.72%, respectively. This work indicates that the proposed fully automatic segmentation method may be accurately and efficiently applied to Gas and Sol muscles segmentation in shank US images.