Multi-directional Spine X-ray Image Segmentation Based On Cyclic Residual Skip Connection Structur

In recent years,scoliosis has attracted widespread attention due to its high incidence,strong harm and increasing tendency to affect younger people.Traditional diagnostic methods for this disease not only have drawbacks such as large errors,long processing times,and low diagnostic efficiency,but also cannot meet the needs of such a large audience.However,with the rapid development of deep learning,research on applying it to the field of spinal medical image segmentation has become a hot topic.This work can not only quickly extract target areas such as the spine and vertebrae,alleviate the burden of doctors in reading films,but also provide convenience for subsequent measurement of spinal parameters and disease diagnosis,promote the development of AI diagnosis and treatment systems for scoliosis,and optimize the diagnosis and treatment of this disease.Meanwhile,X-ray imaging,with its advantages of low cost and small radiation dose,has become the main medical image used in the diagnosis of scoliosis.In addition,diagnosis of this disease requires comprehensive consideration of the patient’s spinal and vertebral morphology from multiple angles.Therefore,based on the theoretical foundation of deep learning,this paper carries out semantic segmentation of the spinal and vertebral regions in multi-angle X-ray images,with specific details described as follows:Firstly,in order to ensure that deep learning models have sufficient data sources for training and testing when carrying out segmentation tasks,we collaborated with a tertiary hospital to collect and construct a spinal X-ray image set containing coronal,sagittal,and left-right bending positions.At the same time,we proposed a two-step spinal automatic segmentation framework.Firstly,the spine and sacrum region in the four-direction images were initially segmented to locate the spinal region in the images.Secondly,combined with image morphological algorithms,fine segmentation of each vertebral body was achieved within the target region.Secondly,a Recurrent Residual Skip Connection Network(RRSC-Net)is proposed for preliminary segmentation of the spine and sacrum area in 4-directional X-ray images.The network employs recurrent residual pathways to compensate for information loss caused by feature fusion at skip connections.The use of multi-scale skip connections and Inception modules enhances the network’s exploration of multi-scale information,while spatial and channel attention modules suppress irrelevant interference.Experimental results show that the network achieves precise segmentation of the spine and sacrum area with IOU values of 92.89%,93.30%,94.10%,and 93.74% in coronal,sagittal,left bending and right bending X-ray images,respectively,which realizes the accurate segmentation of the target.Thirdly,a dynamic vertebral body segmentation algorithm based on the sliding window mechanism was developed and applied to accurately segment multiple-angle X-ray images of the vertebral body region.The sliding window technique was used to slice the vertebral body and narrow the segmentation task from a global to a local view,which improved the network’s perception of the details of the target contour.The use of RRSC-Net resulted in better segmentation performance,and a series of image optimization processes made the predicted results more precise.Experimental results showed that compared with the vertebral body segmentation algorithms of cascade localization FCN and segmentation FCN,the IOU values of vertebral body segmentation in coronal,left bending and right bending position images were increased by 8.32%,5.91% and 13.27% respectively,indicating a significant improvement in segmentation performance.In conclusion,this paper shows in detail the application of deep learning in the scene of spine and vertebral body segmentation in multi-directional X-ray images,which provides important significance for the intelligent and personalized development of clinical diagnosis of scoliosis.