A Fully-Automatic Segmentation of the Carpal Tunnel from Magnetic Resonance Images Based on the Convolutional Neural Network-Based Approach

• Published in: Journal of medical and biological engineering Vol. 41; no. 5; pp. 610 - 625
• Main Authors: Yang, Tai-Hua, Yang, Cheng-Wei, Sun, Yung-Nien, Horng, Ming-Huwi
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.10.2021; Springer Nature B.V
• Subjects: Artificial neural networks; Biological Techniques; Biomedical and Life Sciences; Biomedical Engineering and Bioengineering; Biomedical Engineering/Biotechnology; Biomedicine; Carpal tunnel syndrome; Feature extraction; Feature maps; Image processing; Image segmentation; Magnetic resonance imaging; Median nerve; Neural networks; Original Article; Performance indices; Regenerative Medicine/Tissue Engineering; Tendons; MaskTrack; Carpal tunnel syndrome; Modified DeepLabv3; Segmentation; Convolutional neural networks; Magnetic resonance
• ISSN: 1609-0985; 2199-4757; 2199-4757
• DOI: 10.1007/s40846-021-00615-1

Purpose Carpal tunnel syndrome is one of the common peripheral neuropathies. For magnetic resonance imaging, segmentation of the carpal tunnel and its contents, including flexor tendons and the median nerve for magnetic resonance images is an important issue. In this study, a convolutional neural network (CNN) model, which was modified by the original DeepLabv3 + model to segment three primary structures of the carpal tunnel: the carpal tunnel, flexor tendon, and median nerve. Methods To extract important feature maps for segmentation of the carpal tunnel, flexor tendon, and median nerve, the proposed CNN model termed modified DeepLabv3 + uses DenseNet-121 as a backbone and adds dilated convolution to the original spatial pyramid pooling module. A MaskTrack method was used to refine the segmented results generated by modified DeepLabv3 + , which have a small and blurred appearance. For evaluation of the segmentation results, the average Dice similarity coefficients (ADSC) were used as the performance index. Results Sixteen MR images corresponding to different subjects were obtained from the National Cheng Kung University Hospital. Our proposed modified DeepLabv3 + generated the following ADSCs: 0.928 for carpal tunnel, 0.872 for flexor tendons and 0.785 for the median nerve. The ADSC value of 0.8053 generated the MaskTrack that 0.22 ADSC measure were improved for measuring the median nerve. Conclusions The experimental results showed that the proposed modified DeepLabv3 + model can promote segmentations of the carpal tunnel and its contents. The results are superior to the results generated by original DeepLabv3 + . Additionally, MaskTrack can also effectively refine median nerve segmentations.