基于自适应网格变形的膝关节软骨重建-《中国医学物理学杂志》

文章信息/Info

Title:: Knee cartilage reconstruction based on adaptive mesh deformation

文章编号:: 1005-202X（2026）09-1178-10

作者:: 石鑫恒; 程远志; 青岛科技大学信息科学技术学院，山东青岛 266061

Author(s):: SHI Xinheng; CHENG Yuanzhi; College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China

关键词:: 膝关节软骨重建; 自适应多模板选择; 多模态特征融合; 渐进式网格变形; 图卷积网络

Keywords:: Keywords: knee cartilage reconstruction adaptive multi-template selection multi-modal feature fusion progressive mesh deformation graph convolutional network

分类号:: R318；TP391.41

DOI:: DOI:10.3969/j.issn.1005-202X.2026.09.004

文献标志码:: A

摘要:: 针对膝关节软骨三维重建中个体解剖差异大、薄层结构及局部细节恢复困难等问题，提出基于自适应多模板与多模态特征融合的渐进式网格重建方法AMD-Net。首先，构建基于病情分级的多模板库，通过模板竞争机制自适应选择与患者解剖形态匹配的初始网格；随后，将多尺度体素语义特征与点云几何特征在网格顶点上进行空间对齐和门控融合；最后，利用三阶段图卷积网络实现由全局形态校准到局部细节优化的渐进式重建。在OAI和FastMRI两个公开膝关节MRI数据集上进行实验。结果表明，AMD-Net在OAI数据集上对胫骨、股骨和髌骨软骨的Dice相似性系数分别为86.52%、87.34%和88.57%，在FastMRI数据集上分别为83.42%、86.44%和87.73%；在FastMRI数据集上，股骨软骨的Hausdorff距离为19.14 mm，低于MeshDeformNet的22.34 mm。消融实验中，移除自适应模板初始化后，平均Dice相似性系数由87.48%降至84.99%，平均对称表面距离由1.49 mm增至1.85 mm。上述结果表明该方法能有效融合影像语义与几何结构信息，提高膝关节软骨重建精度和表面拟合质量，其中，自适应模板初始化对重建性能具有重要作用。

Abstract:: Abstract: A progressive mesh reconstruction method termed AMD-Net that combines adaptive multi-template selection with multi-modal feature fusion is developed to address substantial interindividual anatomical variations and overcome the challenges in restoring thin structures and local details during three-dimensional knee cartilage reconstruction. A multi-template library stratified by disease severity is constructed, and an initial mesh matching the patient-specific anatomy is adaptively selected via a template competition mechanism. Subsequently, multi-scale voxel-level semantic features and point cloud geometric features are spatially aligned at mesh vertices and integrated through gated fusion. Finally, a three-stage graph convolutional network is used to progressively refine the mesh from global shape alignment to local detail optimization. Experiments on two public knee MRI datasets, OAI and FastMRI, show that AMD-Net achieves Dice similarity coefficients of 86.52%, 87.34%, and 88.57% for tibial, femoral, and patellar cartilage on the OAI dataset, and 83.42%, 86.44%, and 87.73% on the FastMRI dataset, respectively. On the FastMRI dataset, the Hausdorff distance of femoral cartilage was 19.14 mm, lower than the value of 22.34 mm obtained by MeshDeformNet. The ablation study reveals that removing adaptive template initialization reduces the mean Dice similarity coefficient from 87.48% to 84.99% and increased the average symmetric surface distance from 1.49 mm to 1.85 mm. These findings demonstrate that the AMD-Net can effectively integrate image semantics and geometric information to improve knee cartilage reconstruction accuracy and surface fitting quality, with adaptive template initialization making an important contribution to the reconstruction performance.

《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]

文章信息/Info

备注/Memo

常用功能

导航/Navigate

工具/Tools

统计/Statistics