基于通道残差嵌套U结构的CT影像肺结节分割方法  

CT image segmentation of lung nodules based on channel residual nested U structure

作　　者：蒋武君 支力佳[1,2,3] 张少敏 周涛[1,3] JIANG Wu-jun;ZHI Li-jia;ZHANG Shao-min;ZHOU Tao(School of Computer Science and Engineering,North Minzu University,Yinchuan Ningxia 750021,China;Medical Imaging Center,Ningxia Hui Autonomous Region People's Hospital,Yinchuan Ningxia 750000,China;The Key Laboratory of Images&Graphics Intelligent Processing of State Ethnic Affairs Commission,Yinchuan Ningxia 750021,China)

机构地区：[1]北方民族大学计算机科学与工程学院,宁夏银川750021 [2]宁夏回族自治区人民医院医学影像中心,宁夏银川750000 [3]北方民族大学图像图形智能处理国家民委重点实验室,宁夏银川750021

出　　处：《图学学报》2023年第5期879-889,共11页Journal of Graphics

基　　金：宁夏自然科学基金项目(2021AAC03198,2023AAC03263);国家自然科学基金项目(61561002,62062003);宁夏医学影像临床研究中心创新平台建设项目(2018DPG05006)。

摘　　要：早诊断早治疗对提升肺癌的存活率至关重要。肺结节是肺癌早期主要表现,但其异质性特征增加了计算机断层扫描对肺结节的检测难度,降低了分割结果的精确度。为提高肺结节分割结果的完整性和精确度,提出三维通道残差嵌套U网络(CR U2Net)。浅层特征同时包含病灶细节和噪声信息,提出浅层信息处理U结构平衡噪声信息的干扰;为加强不同层特征信息的交互,丰富特征表达和传递,提出通道残差结构,配合嵌套U结构实现特征信息的提取优化;考虑到浅层特征包含空间细节信息而深层特征具有语义抽象性,设计通道挤压U结构实现不同语义级别特征有效融合;将上述模块集成到UNet中构建出基于嵌套U结构的肺结节分割模型。提出的模型在Lung Image Database Consortium and Image Database Resource Initiative数据集中进行训练,达到了83.83%的Dice系数。优于多数现有肺结节分割方法且与UNet,UNet++以及PCAMNet网络相比领先了3.98%,1.96%和1.26%;针对网络结构进行有效性验证,结果表明各模块均发挥作用,在可接受参数量和计算量的情况下达到最优性能。Early diagnosis and treatment are pivotal in elevating the chances of lung cancer survival.Early-stage lung cancer often manifests through lung nodules.However,their heterogeneity poses a challenge in their detection of lung nodules via computed tomography,subsequently diminishing the accuracy of segmentation results.To improve the completeness and accuracy of lung nodule segmentation results,a 3D channel residual nested U-network(CR U2Net)was proposed for lung nodule segmentation.The shallow information processing U-structure(SIPU)was proposed to address the challenge of managing the interference of noise information while simultaneously incorporating key lesion details within shallow features.To enhance the interaction across different layers of feature information,and to enrich feature expression and transfer,the Channel Residual structure was introduced in conjunction with the nested U-structure to extract and optimize feature information.Acknowledging the spatial detail information found in shallow features and the semantic abstraction in deep features,the channel extrusion U-structure(CEU)was designed to effectively fuse features at different semantic levels.By integrating the proposed modules into UNet,a lung nodule segmentation model based on nested U-structures was constructed.The proposed model was trained on the Lung Image Database Consortium and Image Database Resource Initiative(LIDC-IDRI)dataset.And chieved the best Dice Similarity Coefficient performance,reaching 83.83%.This outperformed UNet,UNet++,and PCAMNet networks by 3.98%,1.96%,and 1.26%,respectively.In addition,ablation experiments were conducted to evaluate the structural validity of the proposed CR U2Net,demonstrating that each module within the proposed segmentation algorithm contributes to achieving optimal performance while adhering to acceptable parameter and computational constraints.

关 键 词：深度神经网络 肺结节分割 通道残差结构 嵌套U结构 通道挤压模块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]