基于边缘约束和改进Swin Unetr的复杂器官分割方法  

Complex Organ Segmentation Based on Edge Constraints and Enhanced Swin Unetr

作　　者：彭琳娜 张红云[1] 苗夺谦[1] PENG Linna;ZHANG Hongyun;MIAO Duoqian(College of Electronic and Information Engineering,Tongji University,Shanghai 201804,China)

机构地区：[1]同济大学电子与信息工程学院,上海201804

出　　处：《计算机科学》2025年第4期177-184,共8页Computer Science

基　　金：国家自然科学基金(62076182,62376198);国家重点研发计划(2022YFB3104700);上海市自然科学基金(22ZR1466700)。

摘　　要：针对腹部CT多器官分割任务中器官边缘模糊、器官比例差异过大的问题,提出了基于边缘约束和改进Swin Unetr的复杂器官分割方法。为了在不同体素比例的器官上提取精细程度不同的特征,设计了掩码注意力模块,通过计算各个器官的掩码信息,提取对应特征。随后,以数据集先验和掩码信息为基础,在相应的窗口和块大小上进行特征提取,以获得小比例器官分割所需的精细化特征,并与编码器的输出特征融合;同时,输出初步预测的语义分割结果后,为了充分利用边界信息,增强模型对于边界信息的处理能力,输出的语义特征通过卷积层进一步提取出边界信息,通过边缘损失最小化使模型的语义分割结果受到边缘预测任务的约束。在BTCV和TCIA pancreas-CT数据集上对所提方法进行训练和测试,在基于卷积网络的UNet++和基于Transformer的Swin Unetr上加入了提出的改进模块并进行训练,与Unetr等经典网络进行了对比实验。在BTCV数据集上,所提模型Dice系数分别达到了0.8479和0.8406,HD距离分别为11.76和8.35,整体上优于其他对比方法,从而验证了所提方法的有效性和可行性。To address the challenges of organ edge blurring and significant differences in organ proportions in abdominal CT multi-organ segmentation tasks,this paper proposes a complex organ segmentation approach based on edge constraints and enhanced Swin Unetr.To extract features of varying degrees of granularity from organs with different voxel proportions,this study introduces the Masked Attention Block.By computing the mask information of each organ,the corresponding features are extracted.Subsequently,based on dataset priors and mask information,refined feature extraction is conducted within appropriate window and block sizes to obtain the fine-grained features necessary for segmenting organs with smaller voxel proportions.Upon generating preliminary semantic segmentation predictions,to fully leverage boundary information and enhance the model’s ability to handle such information,the semantic features are further extracted through convolutional layers to capture boundary details.The model’s semantic segmentation results are constrained by the edge prediction task through edge loss minimization.The proposed method is trained and tested on the BTCV and TCIA pancreas-CT datasets.The enhancement modules are incorporated into the UNet++based on convolutional networks and the Swin Unetr based on Transformers for training.Comparative experiments are conducted with classic networks such as Unetr.On the BTCV dataset,the Dice coefficients reach 0.8479 and 0.8406,with corresponding Hausdorff distances of 11.76 and 8.35,respectively.Overall,the proposed method outperforms other compa-rative methods,confirming its effectiveness and feasibility.

关 键 词：图像分割 注意力机制 计算机断层扫描分割 深度学习 多器官分割 多任务学习 计算机辅助诊断 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]