Abstract

Info

Title:: Automatic segmentation of identified vertebral bones from CT images using CA-SegResNet

Author(s):: ZHU Zhongqi1; GAO Xiaolong2; LI Yinghao1; YANG Guang1; HAO Liguo3; WANG Hongzhi1; 1. Shanghai Key Laboratory of Magnetic Resonance, School of Physics and Electronic Science, East China Normal University, Shanghai 200062, China 2. Department of Imaging, Fujin Chinese Medicine Hospital, Fujin 156100, China 3. Laboratory of Molecular Imaging, School of Medical Technology, Qiqihar Medical University, Qiqihar 161006, China

Keywords:: Keywords: deep learning computed tomography vertebral segmentation segmentation network coordinate attention

Abstract:: Abstract: A three-dimensional (3D) medical image segmentation network (CA-SegResNet) which incorporates a 3D coordinate attention mechanism is proposed to address the issue of segmenting identified vertebral bones from spinal computed tomography (CT) images. The network extracts image features through a deep residual convolutional neural network and fuses the feature maps from each encoder layer with the input of the corresponding decoder layer. Subsequently, a 3D coordinate attention module is introduced to capture inter-channel relationships as well as directional and positional information, establishing long-range dependencies across different spatial directions, thereby enabling precise segmentation of the identified vertebral bones. For the segmentation tasks involving the identified cervical vertebra (the 7th cervical vertebra) and the identified thoracic vertebra (the 12th thoracic vertebra) across 105 cases, CA-SegResNet achieves average Dice similarity coefficients (DSC) of 0.934 5 and 0.918 9 on the test set, with average Hausdorff distances (HD) of 7 and 8 mm. Compared with U-Net results, the average DSC is improved by 0.014 5 and 0.046 3, while average HD is reduced by 176 and 388 mm. The results demonstrate that the network can realize the precise segmentation of identified vertebral bones from CT images.