L3 Paravertebral Muscle Automatic Segmentation System Based On Deep Learning And Its Preliminary Application In Preoperative Evaluation Of Lumbar Spondylolisthesis

First part:Evaluation of the efficacy of the deep learning-based L3 paraspinal automated segmentation and measurement system.Purpose:Exploring the feasibility of developing a deep learning neural network model for automated segmentation and measurement of the L3 paraspinal muscle based on CT data.Materials and Methods:87 patients with lumbar spondylolysis in our hospital(20 males and 67 females with an average age of 55.5±14.3 years).A total of 295 axial CT lumbar spine images through the interlaminar layer of the third lumbar vertebra.Establish a training set:235 pieces(80%)were selected to train and validate a deep learning model developed based on U-net network to segment and measure the functional muslce cross-sectional area(FCSA),intermuscular fat cross-sectional area(IFCSA),while the CT mean value and fat infiltration rate of paraspinal muscle at the same level automatically.Establishing a test set:selecting 60(20%)images,all of which are manually measured and calculated by two physicians(manual group)for functional muscle area,intermuscular fat area,paraspinal muscle CT average value and FIR,and compared with the results of model automatic measurement(model group).Evaluating the model:using the average pixel accuracy,average intersection over union,Dice index,and Hausdorff distance to comprehensively evaluate the model segmentation effect,and using the intraclass correlation coefficient and Bland-Altman method to compare the consistency of the manual group and model group measurement results.Results:he final overall paraspinal muscle MPA,MIoU and Dice of this neural network model are all above 0.90,and the Hausdorff distance is 14.37.In the test set,for the total area,CT average value and FIR of the paraspinal muscle,the ICC values between the manual group and the model group are 0.916,0.971 and 0.953 respectively,showing that the two measurement methods are highly consistent.The results showed that the paraspinal muscle area:the manual group was 74.51 ±19.73 cm2,and the model group was 71.04 ± 17.55 cm2;the paraspinal muscle CT average value:the manual group was 27.01 ± 16.95 Hu,and the model group was 28.03±13.46 Hu;the paraspinal muscle FIR:the manual group was 0.076 ± 0.0437,and the model group was 0.0724 ±0.0458.The Bland-Altman analysis results showed that the two methods were highly consistent.Conclusions:The deep learning neural network model has good consistency with the manual measurement results for the automatic recognition,segmentation and measurement of the L3 paraspinal muscle area,CT average value and fat infiltration rate of CT axial images.The second part:Preliminary application of automatic paravertebral muscle segmentation system in preoperative evaluation of lumbar spondylolisthesis.Purpose:Application of improved U-Net deep learning neural network automatic measurement and segmentation system to automatically measure L3 paraspinal muscles FIR and retrospective analysis of risk factors of single-segment lumbar spondylolisthesis to analyze the efficiency of each factor in predicting single-segment lumbar spondylolisthesisMaterials and Methods:A retrospective analysis of L3 paraspinal muscles CT dataset of 55 single-segment lumbar spondylolisthesis patients(9 males and 46 females)and 51 healthy controls(13 males and 38 females)was conducted,collecting clinical data of two groups and measuring lumbar lordosis(LL)and L3 paraspinal muscles fat infiltration rate(FIR).FIR was obtained by automatic segmentation and measurement of paraspinal muscles on lumbar spine CT axial images based on improved U-Net deep learning neural network system.Binary single factor and multivariate logistic regression analysis were used to determine the independent risk factors of single segment lumbar spondylolisthesis,and receiver operating characteristic(ROC)curve analysis was used to analyze the efficiency of independent risk factors in predicting single-segment lumbar spondylolisthesisResults:The difference between the age,diabetes history,lumbar lordosis(LL)and fat infiltration rate(FIR)of the two groups of patients had statistical differences(P<0.05).The results of binary single factor regression analysis showed that age,diabetes history,lumbar lordosis(LL)and fat infiltration rate(FIR)were all influencing factors of lumbar spondylolisthesis(P<0.05).Binary multivariate logistic regression analysis showed that age,lumbar lordosis(LL)and fat infiltration rate(FIR)were independent risk factors for single-segment lumbar spondylolisthesis.This study explored the efficacy of using LL combined with L3 paraspinal FIR to predict single-segment lumbar spondylolisthesis,with results showing an AUROC of 0.87,a sensitivity of 0.98,and a specificity of 0.65.Conclusions:The results of this study showed that both LL and L3 paraspinal FIR are independent risk factors for single-segment lumbar spondylolisthesis;furthermore,the efficacy of using LL combined with L3 paraspinal FIR to predict single-segment lumbar spondylolisthesis was higher than that of single-factor prediction.