【作者】 王军；

【导师】 赵秀阳；

【作者基本信息】 济南大学 ， 计算机技术（专业学位）， 2025， 硕士

【摘要】 随着人口老龄化的加剧和医疗需求的持续增长,医学图像分割已成为现代医疗体系中不可或缺的关键技术,其通过精准提取医学影像中的目标解剖结构或病灶区域,能够有效辅助医生进行临床诊断并提升诊断效率,具有重要的临床应用价值。然而,现有的医学图像分割方法在临床实际应用中面临两大挑战:一是医学图像标注数据稀缺。由于医学图像的标注过程依赖于专业知识,并且标注成本高昂,可用于训练的标注数据十分有限;二是域偏移问题。由于患者群体、扫描设备及成像参数的差异,不同医疗中心采集的医学图像数据之间存在显著的分布差异,导致模型在跨医疗中心应用时性能大幅下降。因此,为了满足临床应用需求,实现基于少量标注数据且具备跨中心泛化能力的半监督域泛化医学图像分割是至关重要的。半监督域泛化医学图像分割通过利用少量源域标注数据和大量源域无标注数据进行模型训练,能够有效提升模型在不同临床中心数据上的泛化能力。然而,现有算法的性能受限于以下两个问题:(1)医学图像标注数据稀缺和目标域数据未知的问题。(2)不同临床中心的医学图像数据间存在分布差异的问题。围绕上述两个问题,立足于临床场景的实际需求,本文开展了以下工作:(1)解决医学图像标记数据稀缺和目标域数据未知的问题。本文提出了基于自适应感知增强和置信度感知正则化的半监督域泛化医学图像分割网络(APACAR-Net),针对目标域数据不可知问题,设计了一种自适应感知增强模块,该模块通过模型的泛化性能来指导数据增强过程,从而探索潜在的目标域数据分布。同时,利用自适应增强图像进行多视图一致性学习,以减少对标注数据的依赖并提升模型的泛化性能。针对医学图像标注稀缺的问题,提出了置信度感知正则化方法,该方法通过模型的泛化性能来调整伪标签的指导权重,从而优化一致性正则化的过程并充分挖掘无标注数据的有效信息。(2)解决不同临床中心的医学图像数据间存在分布差异的问题。本文在APACAR-Net的基础上提出了一种基于结构保留双归一化和细节结构注意力的半监督域泛化医学图像分割算法(SPDN-DSA),以增强模型在分割阶段对具有跨域不变性的结构特征的学习。该算法针对跨中心数据分布差异问题,提出了结构保留双归一化模块,通过在频域空间中协调结合批量归一化和实例归一化后特征的风格信息,从而减轻不同临床数据之间的分布差异。并且,提出了细节结构注意力模块,通过对低层细节特征进行多尺度学习,并结合通道和空间注意力机制进行处理,从而提高模型对关键细节结构特征的关注能力,进一步提高模型的泛化性能。(3)本文设计并实现了基于半监督域泛化的医学图像分割系统。该系统集成了上述获取的高精度半监督域泛化医学图像分割模型,并具备医学图像读取、医学图像分割、诊断信息管理、结果保存等功能。系统在Window10平台上进行了功能展示和测试,结果表明该系统具有良好的临床适用性,能够有效辅助医生进行临床诊断。更多还原

【Abstract】 With the increasing aging of the population and the continuous growth of medical demand,medical image segmentation has become an indispensable key technology in the modern medical system.By accurately extracting the target anatomical structure or focal area in medical images,it can effectively assist doctors in clinical diagnosis and improve diagnostic efficiency,which has important clinical application value.However,the existing medical image segmentation methods face two major challenges in clinical practice:one is the scarcity of medical image annotation data.Because the process of medical image annotation depends on professional knowledge and the cost of annotation is high,the annotation data that can be used for training is very limited.The second problem is domain migration.Due to the differences in patient population,scanning equipment and imaging parameters,there are significant differences in the distribution of medical image data collected by different medical centers,resulting in a significant decline in the performance of the model when applied across medical centers.Therefore,in order to meet the needs of clinical application,it is very important to realize semi-supervised domain generalized medical image segmentation based on a small amount of labeled data and with cross-center generalization ability.Semi-supervised domain generalization medical image segmentation can effectively improve the generalization ability of the model on the data of different clinical centers by using a small amount of source domain labeled data and a large amount of source domain unlabeled data for model training.However,the performance of the existing algorithms is limited by the following two problems:(1)Scarcity of medical image annotation data and unknown target domain data.(2)There are distribution differences between medical image data in different clinical centers.Focusing on the above two issues and based on the actual needs of clinical scenarios,this thesis carried out the following work:(1)Solve the problem of scarce medical image labeling data and unknown target domain data.In this thesis,a semi-supervised domain generalized medical image segmentation network(APACAR-Net)based on adaptive perception augmentation and confidence-aware regularization is proposed.To solve the problem of data unknowability in the target domain,an adaptive perception augmentation module is designed.The module guides the data augmentation process through the generalization performance of the model,so as to explore the potential data distribution in the target domain.At the same time,adaptive augmented images are used for multi-view consistency learning to reduce the dependence on annotated data and improve the generalization performance of the model.To solve the problem of scarcity of medical image labeling,a confidence-aware regularization method is proposed.This method adjusts the guide weights of pseudo-labels through the generalization performance of the model,so as to optimize the process of consistency regularization and fully mine the effective information of unlabeled data.(2)To solve the problem of distribution difference between medical image data of different clinical centers.Based on APACAR-Net,a semi-supervised domain generalized medical image segmentation algorithm(SPDN-DSA)based on structure-preserving dual normalization and detail structure attention is proposed to enhance the model’s learning of cross-domain invariant structural features in the segmentation stage.In order to solve the problem of cross-center data distribution difference,a structure-preserving dual normalization module is proposed to reduce the distribution difference between different clinical data by coordinating the style information after batch normalization and instance normalization in the frequency domain space.In addition,a detail structure attention module is proposed,which can improve the model’s ability to pay attention to key detail structure features and further improve the model’s generalization performance through multi-scale learning of low-level detail features combined with channel and spatial attention mechanism.(3)We designs and implements a medical image segmentation system based on semi-supervised domain generalization.The system integrates the high-precision semi-supervised generalized medical image segmentation model obtained above,and has the functions of medical image reading,medical image segmentation,diagnosis information management,result saving and so on.The function of the system is demonstrated and tested on Window10platform.The results show that the system has good clinical applicability and can effectively assist doctors in clinical diagnosis.更多还原