Research On Segmentation Method Of Spinal Ct Image Based On Multi-scale Features And Attention Mechanism

Spinal disease is one of the most common diseases in the human body.In recent years,the incidence of spinal diseases has gradually increased and shown a younger trend.The damage condition of spine involves more than 100 diseases of human body.The most recommended method used to diagnose spinal column damage is Computed Tomography（CT）imaging,which has the characteristics of clear images,no image overlap,and high diagnosis rate.Analyzing the images gotten from this method through the computer-aided diagnosis system can help doctors make more accurate and efficient diagnosis and reduce errors caused by manual intervention.However,due to the complex shape of human spine,the similar structure of its neighbors,the spatial relationship between the vertebrae and the surrounding tissues,various pathologies and individual differences,the automatic segmentation of vertebrae becomes more difficult.Therefore,in view of the above-mentioned difficulties in spine segmentation,this paper proposes a network model based on MA-WNet to complete the task of automatic spine CT image segmentation.First of all,use window width and window level adjustment,adaptive local area stretching histogram equalization and bilateral filtering to preprocess two sets of spine CT imaging data sets.Then,this paper designs a spine CT imaging segmentation model based on the MA-WNet network.Aiming at the irregularity and high complexity of the spine shape,a multi-scale feature extraction module is added in the encoding and decoding stages to enhance the feature extraction capabilities of the network.Simultaneously,an attention mechanism is added to reduce the interference of useless information in the segmentation process.Finally,perform the spine CT image segmentation experiment on the MA-WNet network model,each sub-module is verified by ablation experiments,comparing with other deep learning segmentation models though comparatives experiments,test and analyze the obtained results.Experiment result shows,on the CSI 2014 data set,the segmentation result of the method proposed in this paper is 94.53% DSC,and Io U is 90.53%.While on the VerSe 2020 data set,the segmentation result of the method proposed is 91.38% DSC and Io U is 84.22%.Comparing experiments on these two public data sets CSI 2014 and VerSe 2020,it shows that the method proposed in this paper is superior to other algorithms in accuracy,and the edge details and contour information of the image are maintained well,which has certain clinical application value.