基于自注意力的双通路全脊柱X光图像分割模型  

Self-attention based dual pathway network for spine segmentation in X-ray image

作　　者：师文博 杨环 西永明[2] 段文玉 徐同帅 杜钰堃 SHI Wenbo;YANG Huan;XI Yongming;DUAN Wenyu;XU Tongshuai;DU Yukun(College of Computer Science and Technology,Qingdao University,Qingdao 266071,China;Department of Spinal Surgery,Laoshan Branch,the Affiliated Hospital of Qingdao University,Qingdao 266000,China)

机构地区：[1]青岛大学计算机科学技术学院,山东青岛266071 [2]青岛大学附属医院崂山院区脊柱外科,山东青岛266000

出　　处：《中国医学物理学杂志》2022年第11期1385-1392,共8页Chinese Journal of Medical Physics

基　　金：山东省重点研发计划(2019GGX101021);山东省泰山学者项目(ts20190985)。

摘　　要：全脊柱X光图像(包含脊柱、骶骨及髂骨)分割是目前脊椎疾病智能诊断中首要关键的环节。针对U-Net语义分割算法在全脊柱X光图像多区域分割精度较差的问题,提出一种双通道语义分割算法DAU-Net,通过空间通道与语义通道分别学习空间信息特征与图像语义特征,并在解码器端对两类特征进行融合,获取脊柱X光图像中更精准的分割边界。在空间通道中,使用空洞卷积及残差模块扩大视野域并保留更多远端特征信息。此外,将自注意力机制引入语义通道,并设计不同的自注意力编码与自注意力解码模块构建全局关联信息,实现对多个目标骨骼区域语义分割。实验结果表明,DAU-Net能够有效提高脊柱X光图像上的分割精度,相比U-Net、ResU-Net、Attention U-Net、U-Net++,Dice系数分别提高4.00%、1.90%、4.60%、1.19%。The segmentation of X-ray images of the entire spine including the spine,sacrum and iliac bone is the essential step for the intelligent diagnosis of spine diseases.A semantic segmentation network named dual pathway with self-attention for refined U-Net(DAU-Net)is proposed to solve the problem of poor accuracy of U-Net semantic segmentation algorithm in multi-region segmentation in full-spine X-ray image.DAU-Net adopted spatial pathway and semantic pathway to learn spatial information and semantic information separately,and then combines these two types of features at the decoder,thereby obtaining more accurate segmentation boundaries in full-spine X-ray images.In the spatial pathway,dilated convolutions and residual blocks are used to expand the receptive field and capture the long-range dependency feature information.Furthermore,the self-attention mechanism is applied in the semantic pathway,and different self-attention encoders and self-attention decoders are designed to construct global association to achieve the semantic segmentation of multiple target bone regions.The experimental results show that DAU-Net can effectively improve the segmentation accuracy in full-spine X-ray images,and its Dice coeffcience is 4.00%,1.90%,4.60%and 1.19%higher than those of UNet,ResU-Net,Attention U-Net and U-Net++,respectively.

关 键 词：脊柱图像分割 U-Net 语义分割 双通道网络 自注意力机制 

分 类 号：R318[医药卫生—生物医学工程] R816.8[医药卫生—基础医学]