基于特征增强的多分支U-Net肺结节分割

Pulmonary nodule segmentation using multi-branch U-Net based on feature enhancement

针对肺结节尺度差异大、边界纹理不清晰、背景干扰严重导致的肺结节分割不精确的问题,以3D U-Net为基础,引入Transformer结构,设计一种基于特征增强的多分支U-Net肺结节分割算法。Transformer从全局角度提取肺结节及周边组织的结构特征,浅层3D U-Net提取图像纹理特征;利用上述结构特征及纹理特征进行特征增强;多尺度残差块和3D坐标注意力对3D U-Net进行改进,用于提取特征增强后的肺结节多尺度信息,并在3D U-Net解码器基础上,对深层语义信息进行复用,最终实现肺结节分割。在LIDC-IDRI数据集上对该模型进行验证,精确度、敏感度、Dice相似性系数分别达到90.04%、86.64%、88.80%,综合分割性能优于其他算法。

To address the problem of the inaccurate segmentation of pulmonary nodules caused by large scale differences,unclear boundary texture and serious background interference,a multi-branch U-Net based on feature enhancement is designed for pulmonary nodules segmentation.The method uses Transformer to extract structural features of pulmonary nodules and surrounding tissues from a global perspective,and shallow 3D U-Net to extract the texture features.The extracted both structural and texture features are used for feature enhancement.In addition,a multi-scale residual block and 3D coordinate attention module are designed to modify 3D U-Net for obtaining multi-scale information of pulmonary nodules with enhanced features.Based on 3D U-Net decoder,the deep semantic information is reused for accomplishing the segmentation of pulmonary nodules.The verification on LIDC-IDRI dataset shows that the proposed model has a precision,sensitivity and Dice similarity coefficient of 90.04%,86.64%and 88.80%,respectively,exhibiting superior comprehensive segmentation performance as compared with other algorithms.