Deep Learning Model For Automatic Subregional Segmentation And Classification Of Knee Cartilage And Bone Based On MOAKS Semi-quantitative Score

ObjectiveBased on deep learning(DL),an automatic channel was developed for subregional segmentation of knee cartilage and bone and detection of cartilage defect and bone marrow lesions.Materials and MethodsAll the subjects were selected from the Osteoarthritis Initiative(OAI)database.According to the inclusion and exclusion criteria,100 patients(70 training sets,15 validation sets,15 test sets)were randomly included in the segmentation model and the classification model included 1574 cases(1040 training sets,312 validation sets,222 test sets).Using two 3D double echo steady-state(DESS)MRI data,optimized U-Net++network structure using Polarized self-attention(PSA)mechanism(named it PSA-U-Net++),an automatic segmentation model of knee cartilage and bone subregions was established based on MRI osteoarthritis knee Score(MOAKS).The improved MR-Net network was further used to construct an automatic classification model for cartilage defect and bone marrow lesions in each sub-division.A junior radiologist manually scored the classification test dataset to compare the classification deep learning model performance.Dice similarity coefficient(DSC)and volumetric overlap error(VOE)was used to quantify the automatic segmentation accuracy.Sensitivity,specificity and Area under the curve(AUC)were calculated to evaluate the automatic classification performance.ResultIn this study,the automatic segmentation model of knee bone and cartilage subregion constructed by PSA-U-net ++ network and the original U-net++segmentation model in femur,tibia,patella and corresponding cartilage obtained DSC values of 0.985 vs.0.974;0.984 vs.0.971;0.958 vs.0.938;0.898 vs.0.883;0.857 vs.0.841;0.803 vs.0.789,respectively).For the subregional segmentation of articular cartilage and subchondral bone,except for the cartilage in tibia anterior lateral subregion,the DSC of all sub-divisions were above 0.7.The average DSC value of cartilage and subchondral bone subzones were 0.755 and 0.916 respectively.For the binary classification model of cartilage lesions and BMLs,the AUC values of all subregions were above 0.7.The mean AUC of cartilage defect and bone marrow lesions were 0.891 and 0.858 respectively.Compared with the manual score of the primary radiologist,the sensitivity of the classification model was higher than that of the manual classification(cartilage defect:0.829 vs.0.672;bone marrow lesions:0.811 vs.0.676).ConclusionIn this study,the automatic segmentation model of knee cartilage and bone subregion and the automatic classification model of cartilage defect and bone marrow lesions developed by our optimized U-Net++ segmentation network and MR-Net classification network achieved good performance,providing an automatic DL channel for knee MOAKS scoring,which provides a convenient auxiliary tool for early diagnosis and disease surveillance of knee osteoarthritis(KOA).