Research On The Method Of Pedicles Segmentation And Vertebral Rotation Angles Measurement In Radiographs Based On Deep Learning

Scoliosis is a common three-dimensional deformity of the spine.Vertebral axial rotation（VAR）is one of the basic malformations of scoliosis.Clinical measurement of VAR is of great significance in the evaluation of scoliosis.At present,radiograph is the main imaging method in the medical diagnosis of scoliosis.The Drerup method can measure VAR by measuring the offset of the pedicle in the positive X-ray phase relative to the center of the vertebral body,so as to complete a series of evaluations of scoliosis.The pedicle area in the spine radiographs is small and the contrast is not high.The existing radiograph pedicle segmentation method has the problems of low accuracy and low efficiency.For this,an end-to-end approach is proposed.The U-Net neural network based on the attention mechanism is used to segment pedicles in radiographs of the spine.And according to the pedicle contour obtained by network segmentation,the computer was used to mark its center,and then the VAR was measured by Drerup method,so as to reduce the error of manual measurement.The improvement of this paper based on U-Net mainly includes:（1）After each convolution layer in the improved network,a canonical layer is added to make the input of the neural network at each layer maintain the standard normal distribution to avoid network overfitting and thus accelerate network training.（2）Attention gate module is added before each jump connection,which not only improves the accuracy of the network to extract the target area,but also solves the redundancy problem of model jump connection.（3）Before the first sampling,add Selective Kernel Modules（SKM）.SKM can adaptively adjust the size of feature map,which is helpful to capture the spatial correlation between multiple scaling features in the process of convolution operation.（4）Output the network after passing through a Multi-scale Feature Fusion Convolution Block Attention Module（MF-CBAM）.Among them,MF-CBAM is the combination of multi-scale empty convolution and Convolution Block Attention Module（CBAM）.First,the empty convolution operation with different expansion rates is used to capture feature information of different scales,then the feature images containing multiple scales are fused,and finally the fused features are output after using CBAM.CBAM attention by channel module（CAM）and spatial attention module（SAM）,CAM by getting the weight of each channel and analyze the importance of them,then according to the importance to selectively strengthen channel characteristics depend on each other,then the accurate channel characteristics graph input to SAM,SAM through calculating the weight of all the location information in maps,according to the size of the feature weights selectively focus on the important position,to make it stronger correlation between characteristics,so as to achieve more accurate segmentation.The results of segmentation experiments show that the average accuracy of segmentation of radiographs of pedicle using the U-Net network model based on the attention mechanism is 99.15%,the average Dice coefficient is 89.13%,the average accuracy is 88.61%,and the average recall rate is 90.42%,the average Hausdorff distance is 3.8315 pixels,which is better than existing automated segmentation methods.The experimental results of VAR measurement show that the interobserver reliability of computer measurement results is high and the mean absolute error value is small.This method can effectively help doctors to measure VAR more reliably.