【作者】 郭旭；

【导师】 郭晓新；

【作者基本信息】 吉林大学 ， 工程硕士（专业学位）， 2025， 硕士

【摘要】 医学影像分割是疾病诊断、手术规划的关键技术。3D医学影像的标注成本高昂。本研究聚焦于3D医学影像分割中标签获取困难和模型性能提升的问题,深入探索基于半监督学习的方法,提出两种创新网络架构,基于随机双轴展开的Mamba-MT半监督医学图像分割网络(AEMT-MIS)和双向变形渐进融合多阶约束医学图像分割网络(BDPFMC-MIS),有效解决了相关难题。基于随机双轴展开的Mamba-MT半监督医学图像分割网络通过设计Res-Mamba模块并集成至均值教师模型的V-Net编码器中,借助Mamba块捕捉长距离依赖关系,提升模型对医学图像特征的提取与表达能力。将3D原始图像的2D切片沿x与y轴Z字形展开输入Mamba块,打破图像原有顺序,丰富模型输入信息。引入随机开关机制,将展开数据随机分配到Student网络或Teacher网络,通过对不同Mamba编码顺序的分割结果施加一致性约束,增加训练不确定性,提升泛化能力和分割性能。在LA、LITS和Ki TS19等多个公开数据集上的实验结果表明,AEMT-MIS性能优异,仅略逊于全标签V-Net,且相较于基础模型MT有显著性能提升。消融实验验证了模型各关键组件的有效性,证明了随机开关与双轴展开协同工作对提升网络性能的重要性。双向变形渐进融合多阶约束医学图像分割网络针对传统2D配准网络和3D分割网络联合训练时的不足进行创新。在配准网络方面,基于U-Net架构实现源图像到目标图像、目标图像到源图像的双向编码融合,提升变形场计算精度,为生成高质量伪标签奠定基础。对配准网络提出全新的伪标签生成策略,利用双向变形场的信息互补优势,加权融合,提升伪标签质量。设计配准网络和分割网络之间的渐进式伪标签融合策略,根据训练阶段,切片与给定标签的距离动态调整配准和分割网络结果在伪标签中的占比,充分发挥两个网络的优势,实现协同优化。对分割网络Mean Teacher模型的V-Net金字塔解码器多层输出施加一致性约束,挖掘中间层特征,增强模型的鲁棒性和分割准确性。在多个数据集上与前沿半监督方法的对比实验中,该网络在Dice、Jaccard、HD(voxel)和ASD(voxel)等指标上表现出色,分割精度高且与真实标签偏差小。更多还原

【Abstract】 Medical image segmentation is a key technology for disease diagnosis and surgical planning.The annotation cost of 3D medical images is high.This study addresses the challenges of limited label availability and performance enhancement in3D medical image segmentation through semi-supervised learning.I propose two innovative architectures:the Axially-Expanded Mamba-Augmented Mean Teacher for Medical Image Segmentation(AEMT-MIS)and the Bidirectional Deformation Progressive Fusion Multi-order Constrained Medical Image Segmentation Network(BDPFMC-MIS),effectively overcoming annotation constraints while improving segmentation accuracy.For AEMT-MIS,I design the Res-Mamba module and integrate it into the V-Net encoder.By using the Mamba blocks to capture long-range dependencies,it enhances the model’s ability to extract and represent medical image features.3D volumes are processed via zigzag-axis unfolding of 2D slices along orthogonal planes(x/y),enriching spatial context while disrupting slice order,breaking the original order of the images and enriching the input information of the model.A stochastic switching mechanism randomly routes expanded data streams to teacher or student networks,enforcing prediction consistency between divergent expansion orders.By imposing consistency constraints on the segmentation results of different Mamba encoding orders,the training uncertainty is increased,and the generalization ability and segmentation performance are improved.Experimental results on multiple public datasets such as LA,Li TS,and Ki TS19 show that AEMT-MIS has excellent performance,slightly inferior to the fully labeled V-Net,and has a significant performance improvement compared to the basic model MT.Ablation study verify the effectiveness of each key component of the model and prove the importance of the collaborative work of the random switch and biaxial expansion in improving the network performance.The BDPFMC-MIS innovates in response to the deficiencies of traditional registration networks and segmentation networks.In terms of the registration network,based on the U-Net architecture,bidirectional encoding fusion from the source image to the target image and from the target image to the source image is realized,improving the calculation accuracy of the deformation field and laying the foundation for generating high-quality pseudo-labels.A new pseudo-label generation strategy for the registration network is proposed.By taking advantage of the complementary information of the bidirectional deformation fields and performing weighted fusion,the quality of the pseudo-labels is improved.A progressive pseudo-label fusion strategy between the registration network and the segmentation network is designed.According to the training stage and the distance between the slices and the given labels,the proportion of the results of the registration and segmentation networks in the pseudo-labels is dynamically adjusted,giving full play to the advantages of the two networks and achieving collaborative optimization.Consistency constraints are imposed on the multi-layer outputs of the V-Net pyramid decoder of the Mean Teacher model in the segmentation network,mining the features of the middle layers and enhancing the robustness and segmentation accuracy of the model.In comparative experiments with cutting-edge semi-supervised methods on multiple datasets,this network performs excellently in terms of indicators such as Dice,Jaccard,HD(voxel),and ASD(voxel),with high segmentation accuracy and a small deviation from the true labels.更多还原