Research on Cervical Vertebrae Segmentation Based on Improved UNet Network

Cervical spine injury detection heavily relies on the accurate segmentation of cervical vertebrae. However, traditional segmentation methods only provide limited semantic information and exhibit poor segmentation accuracy. To overcome these issues, a modified UNet network-based algorithm was investigated for more precise and detailed cervical vertebral body segmentation.Firstly,the UNet network is incorporated with the residual structure, and optimize the 3x3 convolution used in the down-sampling process to work efficiently with the residual structure.Secondly, the maximum pooling layer and average pooling layer of Channel Attention (CA) mechanism is modified to 1x1 convolutions,and these layers are then integrated into the feature fusion process of the UNet network. The improved UNet network demonstrated better segmentation accuracy compared to other models, with performance improvements of 3.81%, 3.09%, and 1.16% when compared to UNet, Attention-UNet, and HRNet, respectively. The experimental results showed that the improved UNet network improved the segmentation performance of cervical vertebrae.