改进UNet++的食道癌肿瘤分割方法

Segmentation method of esophageal cancer based on improved UNet++

晚期食道癌患者五年生存率仅约20%,准确地检出食道癌肿瘤区域对后续治疗具有重要的临床意义。该文针对当前主流的食道癌病灶分割方法在复杂曲度位置细节丢失、分割精度低等问题,基于UNet++设计了一种改进的网络模型。在编码器中引入SE模块,使网络重点关注待分割区域;将网络关键层X^(i,0)中常规方阵卷积替换为可变形卷积,使网络更好地适应癌灶边界复杂的曲度变化;使用多尺度特征融合,充分提取出肿瘤边界的细微特征;将Encode-Decoder结构升级为双向O型循环结构,来提高网络对特征的使用效率。实验表明,该算法在食道癌CT影像分割任务中,Dice相似性系数可以达到84.81%,相比UNet++提高了6.37%。文中所提出的方法相比现阶段其他先进算法,能更为准确地分割出食道癌肿瘤的不规则边界。

The 5-year survival rate of patients with advanced esophageal cancer is only about 20%.Accurate detection of esophageal cancer region is of great clinical significance for follow⁃up treatment.In this paper,an improved network model based on UNet++is proposed to solve the problems of missing details of complex curvature position and low segmentation accuracy.The SE module is introduced into the encoder to make the network focus on the region to be divided,the normal square matrix convolution is replaced by deformable convolution in X^(i,0)to make the network more adaptable to the complex curvature of the tumor boundary,the multi⁃scale feature fusion is used to fully extract the subtle features of the tumor boundary,and the Encode⁃Decoder structure is upgraded to a bidirectional O⁃type cyclic structure to improve the efficiency of the network to use the features.Experimental results show that the Dice similarity coefficient can reach 84.81%in CT image segmentation of esophageal cancer,which is 6.37%higher than UNet++.Compared with other advanced algorithms,the method proposed in this paper can segment the irregular boundary of esophageal cancer accurately.