结构先验引导的多模态腰椎MRI图像分割算法  

Multimodal lumbar MRI image segmentation algorithm guided by structure priori

作　　者：李伟豪 王苹苹 许万博 魏本征[1,2,3] LI Weihao;WANG Pingping;XU Wanbo;WEI Benzheng(Center for Medical Artificial Intelligence,Shandong University of Traditional Chinese Medicine,Qingdao 266112,Shandong,China;Qingdao Academy of Chinese Medical Science,Shandong University of Traditional Chinese Medicine,Qingdao 266112,Shandong,China;Qingdao Key Laboratory of Artificial Intelligence Technology of Traditional Chinese Medicine,Qingdao 266112,Shandong,China;Qilu Hospital of Shandong University Dezhou Hospital,Dezhou 253046,Shandong,China)

机构地区：[1]山东中医药大学医学人工智能研究中心,山东青岛266112 [2]山东中医药大学青岛中医药科学院,山东青岛266112 [3]青岛市中医人工智能技术重点实验室,山东青岛266112 [4]山东大学齐鲁医院德州医院,山东德州253046

出　　处：《山东大学学报(工学版)》2025年第1期66-76,共11页Journal of Shandong University(Engineering Science)

基　　金：国家自然科学基金资助项目(62372280,61872225);山东省自然科学基金资助项目(ZR2020KF013,ZR2019ZD04);青岛市科技惠民示范专项资助项目(23-2-8-smjk-2-nsh);齐鲁卫生与健康领军人才资助项目。

摘　　要：为挖掘腰椎磁共振成像(magnetic resonance image,MRI)图像中多种模态信息的相关性、腰椎结构间的相互依赖关系、腰椎结构先验知识对腰椎精准分割和疾病辅助诊断的重要价值,提出一种结构先验引导的多模态信息融合分割算法。设计的多模态图像编码模块(multi-modality encoding module,MMEM)可同时对T1和T2加权图像做语义特征提取;跨模态体素融合模块(cross-modality voxel fusion module,CMVF)可在融合过程中为各模态图像特征自适应分配融合权重。根据腰椎内部各组织结构间的先验知识构建图模型,利用图卷积神经网络分割模块(graph convolutional networks segmentation module,GCNSM)实现图模型上的语义信息传播。采用多模态图像解码模块(multi-modality decoding module,MMDM)对特征图进行解码,实现对椎体及椎间盘的精准图像分割。对山东大学齐鲁医院德州医院采集的190组患者MRI数据集进行试验验证,所设计算法的平均骰子系数Dice、交并比IoU、95%Hausdorff距离HD95和平均对称表面距离ASSD分别为90.3%、82.31%、4.40 mm和1.21 mm,结果表明了算法的有效性及先进性。To explore the relevance of multi-modality information in lumbar magnetic resonance image(MRI)images,the interdependence among lumbar structures,and the significant value of prior knowledge of lumbar structures for precise segmentation of lumbar spine and disease-assisted diagnosis,a segmentation algorithm guided by structural priori for multi-modality information fusion was proposed.A multi-modality encoding module(MMEM)was designed,which could simultaneously extract semantic features from T1-weighted and T2-weighted images.A cross-modality voxel fusion module(CMVF)could adaptively allocate fusion weights to the features of each modality image during the fusion process.The graph model was constructed according to the prior knowledge of the internal organizational structures of the lumbar,and the graph convolutional networks segmentation module(GCNSM)was utilized to propagate semantic information on the graph model.Multi-modality decoding module(MMDM)was employed to decode the feature maps,which achieved precise image segmentation of the vertebrae and intervertebral discs.The MRI data set of 190 groups of patients collected from the Qilu Hospital of Shandong University Dezhou Hospital was verified by experiments.The average dice coefficient Dice,intersection of union IoU,95%Hausdorff distance HD95,and average symmetric surface distance ASSD of the designed algorithm were 90.3%,82.31%,4.40 mm,and 1.21 mm,respectively,demonstrated the effectiveness and advancement of the algorithm.

关 键 词：腰椎 磁共振成像 多模态 多类别分割 图卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]