基于循环残差跳连结构的多方位脊柱X光影像分割

Multi-directional spine X-ray image segmentation algorithm based on recurrent residual jump connection structure

脊柱X光影像分割为脊柱配准、参数测量与疾病分型奠定了基础,提出了一种基于深度学习的多方位脊柱X光影像自动分割方法。该方法级联粗分割网络与细分割网络,采用Inception模块进行特征提取,通过多尺度跳跃连接结构实现二者互连,构建循环残差路径解决跳连处特征融合时易造成信息丢失现象,同时细分割网络瓶颈处引入并行空间和通道挤压激励模块(SCSE)提高网络对脊柱部位的关注。实验结果表明,该方法在冠状位和左、右Bending位X光影像数据集中平均交并比(IoU)为92.89%、94.10%和93.74%,其三方位脊柱分割效果优于全卷积网络(FCN)、U-Net、SegNet等模型。

Spine X-ray image segmentation lays the foundation for spine registration,parameter measurement and disease classification.An automatic multi-directional spine X-ray image segmentation method based on deep learning was proposed.In this method,the rough segmentation network and the fine segmentation network were cascaded,the Inception block was used for feature extraction,the interconnection between them was realized through the multi-scale skip connection structure,the recurrent residual path was built to solve the information loss caused by the feature fusion at the skip connection,and the concurrent Spatial and Channel Squeeze&Excitation module(SCSE)was introduced at the bottleneck of the fine segmentation network to improve the network’s attention to the spine.Experimental results show that the average Intersection over Union(IoU)values of this method in coronal,left bending and right bending X-ray image data sets are 92.89%,94.10%and 93.74%,and its three-directional spine segmentation effect is better than that of Fully Convolutional Network(FCN),U-Net,SegNet and other models.