Research On 3D Spinal Image Segmentation Based On Deep Learning

Spinal segmentation is the key to the diagnosis of spinal diseases,and it is usually done by experienced medical experts to complete purely manual marking,which is very time-consuming and energy consuming.Therefore,an automatic spine image segmentation method is urgently needed.The early traditional methods and machine learn-based methods can not meet the needs of current medical development.In recent years,many spinal image segmentation methods based on deep learning have shown good segmentation effect on healthy spine or local spine,but they are difficult to apply to complex scenes and have poor robustness.In order to improve the segmentation accuracy and robustness,this paper proposes a three-stage spine segmentation method based on deep learning to segment the spine in 3D spine images.The three stages are spine segmentation,vertebrae localization and recognition,and vertebrae segmentation.The specific work is as follows:(1)Aiming at the problem that many spinal columns in 3D spine images are not in the center of the image,this paper proposes an improved 3D-UNet spine segmentation method based on the attention mechanism.By adding AG module at each jump connection of 3D-UNet codec,the irrelevant features are inhibited and the sensitivity of the model is improved.This allows better separation of the spine and soft tissue background in the image.(2)In view of the problem that the adjacent vertebrae in the spine are overlapping and highly similar,which is easy to cause semantic confusion in the segmentation network,the spatial configuration network is used to locate and identify the vertebrae before the vertebrae segmentation.The local appearance components and spatial configuration components in the spatial configuration network coordinate with each other to generate the vertebrae centroid heat map,which can effectively predict the centroid coordinates of each vertebrae and its corresponding vertebrae category label.(3)This paper presents a study based on the residual improved 3D-UNet to complete the final vertebrae segmentation,this improved model in 3D-UNet codec introduced in each phase of the residual module,improve the segmentation performance of network and to prevent the occurrence of gradient model training disappear phenomenon,make model fast convergence in the training process,reduce the training time.(4)Performance analysis of the proposed method is performed on three public data sets.First of all,the Dice coefficient of the proposed method in MR Lower Spine and x Vertseg data sets reaches more than 95%,and the Dice coefficient in Verse2020 data sets reaches 91.62%.Then,in the Verse2020 dataset,the robustness analysis is carried out for different spinal regions,complex scenes containing abnormal vertebrae and noise.The final experimental results show that the proposed three-stage spinal segmentation method has good generalization and high robustness.