双重降维通道注意力门控U-Net的胰腺CT分割  

Dual Dimension Reduction and Channel Attention Gate U-Shaped Network for Pancreatic CT Segmentation

作　　者：纪建兵 陈纾 杨媛媛[3] Ji Jianbing;Chen Shu;Yang Yuanyuan(College of Information Engineering,Fujian Business University,Fuzhou 350012,China;Faculty of Innovation and Design,City University of Macao,Macao 999078,China;Department General of Surgery,Fujian Medical University Union Hospital,Fuzhou 350001,China)

机构地区：[1]福建商学院信息工程学院,福州350012 [2]澳门城市大学创新设计学院,中国澳门999078 [3]福建医科大学附属协和医院普外科,福州350001

出　　处：《中国生物医学工程学报》2023年第3期281-288,共8页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(82001895);福建省自然科学基金(2018Y9039);福建商学院科研创新团队支持计划资助(商院科研〔2021〕72号)。

摘　　要：从腹部CT图像中分割并重建胰腺3D模型对于辅助疾病诊疗有重要意义。由于胰腺在图像中占比小且与周边组织难以区分等原因,现有方法准确性和稳定性不足。本研究提出一种双重降维和通道注意力门控U型网络,在编码路径中以双重降维模块加强浅层特征空间有效信息提取,在编解码连接中嵌入通道注意力门控模块从通道层级过滤冗余特征。在NIH发布的胰腺分割公开数据集上(包括82例CT图像)进行实验,采用集合相似度(DSC)、召回率(R)和精确率(P)验证分割性能,使用三维顶点距离误差(VDE)评估3D重建效果。DSC、R和P值分别达到82.35%±5.76%、81.07%±8.50%、84.04%±5.40%,VDE降低至1.27±0.90,优于U-Net和Attention-Unet等方法。结果表明,所提出方法能够提高胰腺CT图像分割性能,重建的3D模型能够更好反映个体胰腺实际情况。Segmentation and reconstruction of pancreatic 3D model from CT images is of great significance for assisting diagnosis of the disease.Due to the small proportion of pancreas and the difficulty to distinguish pancreas from the surrounding tissues,existing methods are not accurate and stable enough.Herein we proposed a dual dimension reduction(DDR)and channel attention gate(CAG)U-shaped network.The DDR was added to the coding path to strengthen the effective information extraction in the shallow feature space,and the CAG was embedded in the skip connection to filter redundant feature information from channel level.On the public data set of pancreas segmentation published by NIH(including 82 CT samples),we evaluated the segmentation performance by dice similarity coefficient(DSC),recall(R)and precision(P),and evaluated 3D reconstruction through vertex distance error(VDE).DSC,R and P indexes reached(82.35±5.76)%,(81.07±8.50)%,and(84.04±5.40)%respectively,and VDE decreased to 1.27±0.90,which was better than that obtained from U-Net,Attention-Unet and other methods in the experiment.The experimental results showed that the method we proposed improved the segmentation performance of pancreatic CT images,and the 3D reconstruction model reflected the actual situation of individual pancreas.

关 键 词：图像分割 三维重建 胰腺 特征降维 注意力 

分 类 号：R318[医药卫生—生物医学工程]