基于Concat⁃UNet的食管癌肿瘤医学影像分割研究  

Research on Medical Image Segmentation for Esophageal Cancer Tumors Based on Concat-UNet

作　　者：刘文[1] 亓文霞 仲国强 王佳佳 王大寒 LIU Wen;QI Wenxia;ZHONG Guoqiang;WANG Jiajia;WANG Dahan(Faculty of Information Science and Engineering,Ocean University of China,Qingdao,Shandong 266100,China;Department of Oncology,Dezhou People's Hospital,Dezhou,Shandong 253000,China;College of Computer and Information Engineering,Xiamen University of Technology,Xiamen,Fujian 361024,China)

机构地区：[1]中国海洋大学信息科学与工程学部,山东青岛266100 [2]德州市人民医院肿瘤科,山东德州253000 [3]厦门理工学院计算机与信息工程学院,福建厦门361024

出　　处：《计算机工程》2022年第12期312-320,共9页Computer Engineering

基　　金：科技创新2030“新一代人工智能”重大项目(2018AAA0100400);山东省自然科学基金(ZR2020MF131);福建省医疗数据挖掘与应用工程技术研究中心开放课题(MDM2018007);青岛市科技计划项目(21-1-4-ny-19-nsh)。

摘　　要：食管癌肿瘤的诊断方式主要是医生对胸部计算机断层扫描(CT)影像进行阅片。由于医生的主观判断易受外界环境的干扰,因此诊断结果与实际结果存在偏差。基于深度学习的图像分割网络对辅助诊断食管癌肿瘤具有重要意义。因食管在整体胸部CT影像中所占的区域较小且对比度较低,传统的图像分割网络难以准确地确定食管癌肿瘤的区域。为精准分割医学CT影像中的食管癌肿瘤,提出图像分割网络Concat-UNet。基于U-Net网络,采用编码解码模式的U型对称架构对网络中的卷积模块进行改进,并引入跳跃连接和批量归一化层,将卷积模块的原始输入与提取特征后的输出进行特征融合,以增强网络的特征提取能力。在此基础上,采用BCEWithLogits与Dice损失函数相结合的方式联合训练网络。实验结果表明,相比SegNet、ERFNet、U-Net等网络,Concat-UNet在食管癌数据集上的检测精确率为91.87%,相比基准网络U-Net提升了11.64个百分点,具有较优的分割效果。The diagnosis of esophageal cancer tumors mainly depends on the observation of chest images by doctors based on Computed Tomography(CT)scan.Because the doctor's subjective judgment is easily influenced by environmental factors,a deviation may exist between the diagnosis and actual results.The image segmentation network based on deep learning is crucial for the auxiliary diagnosis of esophageal cancer.Because of the small area and low contrast of the esophagus in the entire chest CT image,it is difficult to accurately determine the area of esophageal cancer tumor based on the traditional image segmentation network.This paper proposes an image segmentation network Concat-UNet to accurately segment esophageal cancer tumors in medical CT images.Based on the U-Net network,the convolution module in the network is improved using the U-shaped symmetric architecture of encoding and decoding mode.The skip connection and Batch Normalization(BN)layer are introduced to fuse the original input of the convolution module with the output after feature extraction to enhance the feature extraction capability of the network.On this basis,the combination of BCEWithLogits and Dice loss functions is used to jointly guide the network training to stabilize the network training.The experimental results show that compared with SegNet,ERFNet,U-Net,and other networks,Concat-UNet has a detection accuracy of 91.87%for the esophageal cancer dataset and exhibits an improved segmentation effect.Specifically,Concat-UNet is improved by 11.64 percentage points over the baseline U-Net network.

关 键 词：食管癌 语义分割 U-Net网络 Concat-UNet网络 跳跃连接 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]