| Scoliosis has an incidence rate of about 2-3% in children aged 10-16 years,and it is a disease with a high degree of harm and incidence of spinal diseases.The traditional medical method for diagnosing scoliosis is time-consuming and has large errors.Therefore,the research on an automated scoliosis diagnosis system is very important,and spine segmentation is an important rule of the spine disease diagnosis system.The X-ray image is still one of the important data in the diagnosis of spinal diseases due to their low price,low radiation dose,and cheap equipment.Therefore,this thesis will perform semantic segmentation of spine X-ray images to obtain key information for the diagnosis of the spine,sacrum,and ilium.In this thesis,deep learning algorithms based on channel attention and self-attention mechanism respectively perform semantic segmentation of spine X-ray images and improve the problem of poor X-ray image segmentation by using a dual-channel semantic segmentation framework,and finally achieve better results than other Segmentation algorithms for common models.Aiming at the problem that the direct segmentation effect of the model is poor,this paper constructs a multi-step segmentation framework for spine and vertebral body segmentation,which can still achieve good results for the spine in the small dataset.The specific work of this thesis is described as follows:(1)This thesis proposes a two-step segmentation framework for automatic segmentation of spine X-ray images,which can first locate the spine region(including the spine backbone,sacrum and ilium regions)in the rough stage.Second,the image morphological process operations are used to obtain 18 spine image slices and obtain an accurate spine vertebral body segmentation map in the second segmentation process.(2)To improve the problem of blurred edges in whole spine X-ray image segmentation,this thesis proposes a dual pathway semantic segmentation model AGNet based on the reverse attention mechanism and the channel attention mechanism.The model uses the semantic information subnetwork and edge information subnetwork to learn different feature information respectively.And the dual-channel features are fused through the edge information feature fusion module which is proposed in this paper to improve the segmentation accuracy of the spine backbone and the vertebral body.Experiments show that compared with the U-Net model,the proposed AGNet in this thesis has about a 3.2% improvement in spine segmentation and a 2.3% improvement in the spine vertebral dataset.(3)To solve the issue of the poor segmentation accuracy of spine X-ray images,this thesis proposes another dual pathway semantic segmentation framework,named DAU-Net which is based on the spatial self-attention mechanism.The DAU-Net uses the self-attention mechanism to construct the association of global spatial location information in image feature maps.The performance consumption of the self-attention mechanism is reduced by downsampling operation.In addition,a dual pathway network is designed to learn image semantic information and image spatial location information respectively.Experiments show that compared with the U-Net model,the proposed DAU-Net in this thesis has about a 4.2% improvement in the spine segmentation dataset. |
