基于膝关节MR图像的分割模型构建及验证

Construction and verification of segmentation model based on knee joint MR image

目的:构建并验证一种膝关节MR图像分割算法,旨在解决软骨细小结构识别困难、分割边界模糊、错分等问题,发现软骨的早期病变,帮助医生诊断膝骨关节炎等慢性疾病。方法:使用膝关节公开数据集SKI10进行实验验证,划分为训练集(60%)、验证集(20%)、测试集(20%);基于Transformer方法和U-Net方法,提出融合通道注意力机制和边界注意力机制的新型网络架构CE-TransUNet;以平均Dice相似系数(DSC)为主要评价指标,探索模型在膝关节MR图像分割中的性能。结果:与经典算法进行对比,CE-TransUNet具有更好的分割效果,其DSC指数达到了90.48%,在股骨和胫骨上DSC分别达到了93.55%和93.10%,在股骨软骨和胫骨软骨上DSC分别达到了87.69%和87.58%。结论:CE-TransUNet分割结果与人工分割结果有很好的一致性。其分割效果优于对比网络模型,为膝关节软骨的自动分割提供了一种新思路,能够帮助临床诊断,有较好的应用前景。

Objective:Develop and validate a knee joint MR image segmentation algorithm to address challenges in accurately identifying the fine structure of cartilage,resolving fuzzy segmentation boundaries,and mitigating mis-segmentation.The goal is to detect early cartilage lesions and aid the diagnosis of chronic diseases such as knee osteoarthritis.Methods:Utilized the SKI10 public dataset for experimental verification,partitioned into training(60%),validation(20%),and test(20%)datasets.CE-TransUNet,a novel network architecture combining the Transformer and U-Net methods,was proposed.This model integrates channel attention and edge attention mechanisms.The performance of the proposed model in knee joint MR image segmentation was assessed using the average Dice similarity coefficient(DSC)as the primary evaluation metric.Results:Compared to the classical algorithm,CE-TransUNet demonstrates superior segmentation performance,achieving a DSC of 90.48%.Specifically,the DSC for femur and tibia segmentation reaches 93.55%and 93.10%,respectively.While for femoral and tibial cartilage,it is 87.69%and 87.58%.Conclusion:The segmentation results obtained using CE-TransUNet closely align with manual segmentation results,indicating superior performance compared to the comparison network model.This method presents a novel strategy to automatic knee cartilage segmentation,holding potential for clinical diagnosis and application.