Construction and verification of segmentation model based on knee joint MR image
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.
medical image segmentationknee joint cartilage segmentationMR imageTransUNetattention mechanism
万方数据
维普
