Precise Prediction Model Of Biomechanical Parameters Of Lower Limbs Based On Statistical Shape Modeling And Deep Learning

Objective:Total knee arthroplasty(TKA)and high tibial osteotomy(HTO)are common methods to reconstruct joint,correct deformity and restore biomechanical axis.Accurate preoperative measurement of BMP and good intraoperative correction are considered to be one of the key factors to determine the success or failure of surgery and long-term efficacy.The BMP collection system based on radiology has been formed,but it has higher requirements on the imaging posture.Compared with the recognized standard posture(SP),the BMP collected by non-standard posture(NSP)deviates from the real value to varying degrees.However,there are few studies on the degree of deviation from the true value and no corrective measures,and manual measurement of BMP is time-consuming,laborious,tedious,and thin data.The existing BMP collection system cannot meet the precision requirements of modern orthopedics.Therefore,this study:(1)MD-MP model spectrum was constructed based on statistical shape modeling(SSM)method,and bmp-bias algorithm model was developed to achieve BMP correction by MD-MP measurement within the research scope;(2)Developing full-length coronal X-ray multi-task model(FLCR-MTM)of lower limbs based on Deep Learning(DL)to achieve accurate prediction of feature points,attitude classification and BMP of lower limbs.Methods:The full-length CT data of the lower limbs of healthy volunteers(n=30)were collected,and the original 3D model of the lower limbs was constructed through 3D reconstruction and data annotation.The computer automation program was developed to generate MD-MP model spectrum of the lower limbs within the research scope,and the BMP of all samples in the model spectrum was calculated and the MD-MP-BMP database was generated.Calculate the BMP difference between all NSP and SP in MD-MP-BMP database and generate BMP-bias database.The correlation,clustering characteristics,equilibrium characteristics and distribution shape of data in BMP-BIAS database were analyzed.Based on BMP-BIAS data,BMP-Bisa algorithm model is developed by machine learning method,and MAE and MSE of the model are tested.The performance of the model in diagnosis consistency was evaluated by criterion-related Validity verification.A lower limb full-length X-ray multitask model(FLCR-MTM)is developed based on U-Net neural network.F-LCR data from 3 research centers were collected retrospectively(n=1400),and independent F-LCR data(n=86)and full-length CT data of lower limbs were collected prospectively(n=86).F-LCR data(n= 1400)were manually labeled with attitude 7 classification label and BMP measurement label.F-LCR was randomly divided into training set(n= 820),verification set(n=290)and test set(n=290)in a ratio of 3:1:1.The FLCR-MTM is trained with the training set and the performance of the model is verified in the verification set.The mean error(MRE mm)of model feature point prediction was tested in the test set,and the success detection rates(SDR %)of 2mm,4mm and 6mm were tested.In the test set,the AUC,specificity and sensitivity of attitude classification were calculated using the area under the subject operating characteristic curve(AUC)method,and the total accuracy of model attitude classification was calculated using confusion matrix.Pearson correlation analysis and Bland-Altman analysis were used to evaluate the diagnostic effect of the two groups and estimate the difference.The performance of FLCR-MTM was evaluated by mean square error(MSE),mean absolute error(MAE)and consistency correlation coefficient(CCC).The two correlations are compared by testing the correlation values by converting them to z-scores.Results:The full-length CT of lower limb of 30 healthy subjects was included and the original 3D annotated model of lower limb was constructed.The MD-MP model spectrum of lower limb developed based on statistical shape modeling method contained 111,600 basic shapes,and the generated MD-MP-BMP database contained1,562,400 samples.The generated BMP-Bias database contains 1,562,400 sample sizes.Analysis of bmp-bias data showed that:(1)MD-MP was significantly correlated with bmp-bias;(2)BMP-BIAS data program smooth slip linear characteristics,linear model with general regularity;(3)The distribution of BMP-BIAS data is uneven,in which HKA,FTA and MPTA BMP-Bias have a large range of changes(between-23.64°～23.84°,-25.21°～24.38°,-28.45°～29.99°).However,FM-FS,JLCA,m LDFA and a LDFA BMP-Bias have a small range of fluctuation(during the period of volatility 2.59°～-0.79°,-3.07°～1.25°,-0.51°～2.59°,-2.45°～1.35°).The MAE and MSE of seven BMP are predicted to be 0.046° and 0.046° respectively 0.083°.FLCR-MTM was developed based on U-NET convolutional neural network,and1400 F-LCR were included in the FLCR-MTM development study.FLCR-MTM was tested in the F-LCR test set(n=290):(1)Point prediction model results show that the total MRE of expert and model feature points are 2.66 mm and 3.08 mm respectively,and the SDR(%)of 2mm,4mm and 6mm in the two groups are 87.50%,96.50%,98.50% and 85.25%,95.25%,97.50%,respectively.(2)580 F-LCR(1160 lower limbs)test sets were used to test the classification model,The AUC of standard posture,mild pronation posture,moderate pronation posture,severe pronation posture,mild pronation posture,moderate pronation posture and severe pronation posture were 0.978,0.978,0.961,0.952,0.959,0.959,respectively,and the overall accuracy of the model was 93.45%(1084: 1160).(3)86 lower limb CT and F-LCR were included in the calibration correlation validation study.The total measurement consistency between F-LCR group and CT group,FLCR-MTM group and CT group was R 0.838(P < 0.01),MAE 6.053(°),MSE 19.192(°),CCC 0.829(95%CI,0.750-0.885),R 0.994(P < 0.01),MAE 1.037(°),MSE 4.053(°),CCC 0.993(95%CI,0.990-0.996).The consistency difference between FLCR-MTM group and F-LCR group was 15.445(P < 0.01).Conclusions:(1)3D reconstruction and standardization of lower limb CT data of healthy people can realize statistical shape modeling of lower limb MD-MP.The MD-MP model spectrum constructed contains the BMP change rule between standard pose shape and non-standard pose shape.(2)Based on the limb MD-MP model spectrum,the BMP-BIAS algorithm model developed can correct the BMP Bias value of MD-MP within the research range.(3)The developed FLCR-MTM can predict bone feature points of lower limbs and classify posture of lower limbs,and achieve accurate prediction of MD-MP-BMP on F-LCR.(4)Compared with traditional BPM collection methods,FLCR-MTM provides corrected BMP to the clinic,which may help improve the surgical treatment level of TKA and THO.