| 【Backgrounds】Adolescent idiopathic scoliosis(AIS)is a kind of spinal deformity which has unknown reasons and severely affects the health of teenagers.Surgical correction with instruments is the main approach to treat middle and severe AIS patients.With the progress of operation concept and equipment,the rate of deformity correction gradually increases.However,due to individual differences and lack of comprehension of spinal compensatory characteristics,patients often accompany with spinal imbalance and distal adding-on phenomenon,which seriously affect the surgical efficacy.Also,the theoretical mechanism and risk factors of such complications have not been unified.As for individuals,the existing AIS classification,risk factor analysis and a great deal of prediction formulas can merely give the general correction principles and basis,but lack personalized and simulated correction strategy selection.At present,with the continuous development of artificial intelligence(AI),deep learning(DL)has injected fresh vitality into the field of deformity,which has been demonstrated in the dimensions of screening and diagnosis,Cobb angle calculation and classification.However,there is still no suitable model for the selection of pedicle screw placement strategy for posterior approach of AIS.【Objectives】This study mainly aimed at the coronary complications of AIS patients,taking the most common adding-on phenomenon in Lenke 1&2 type AIS and the most common immediate postoperative coronary imbalance phenomenon in Lenke 5&6 type AIS as examples,to study the corresponding mechanism and risk factors through the mathematical statistical analysis of clinical data.Meanwhile,combined with different deep learning model,we also aimed to predict the coronal results of AIS posterior orthopedic surgery.And the optimized model was selected according to the prediction results,so as to realize the preliminary exploration of suitable orthopedic model construction.【Methods】1.Characteristics analysis of spontaneous lumbar curvature compensation after selective fusion in Lenke 1&2 AIS patientsFifty-one Lenke 1&2 AIS patients whose lower instrumented vertebrae were L1 were collected.The demographic data and corresponding coronal parameters were analyzed before,immediately after and at the last follow-up.At the same time,the wedge angle of the intervertebral disc of each unfused segment was measured,and the compensatory ability was calculated immediately after surgery and at the last follow-up.The characteristics of local compensatory ability and its relationship with global compensatory ability were analyzed.2.Analysis of risk factors for immediate postoperative coronal imbalance in Lenke 5 & 6 AIS patientsThe data of eighty-five patients with Lenke 5&6 AIS were collected and the full-length spine films were evaluated before,immediately and at the final follow-up.The patients were divided into two groups: the immediate postoperative imbalance(IPCIB)group and the nonIPCIB group.The risk factors leading to IPCIB were analyzed,and the IPCIB index was proposed and verified.3.Preliminary exploration of the construction of adolescent idiopathic scoliosis posterior surgical correction model based on deep learning425 AIS patients who underwent posterior pedicle screw fixation in our hospital and had more than 2 years follow-up were collected.The data were preprocessed by parameterization of image data.Different types of data were converted to the universal continuous high dimensional feature space.The multi-layer depth network model,encoder decoder model,convolution and attention mechanism model and deep factorization machine model were adopted and assessed by the mean square error and mean absolute error between prediction of postoperative spinal coordinates and corresponding real postoperative coordinates in test set,as well as curve simulation results.【Results】In the represented Lenke 1&2 AIS population,41 cases were female(80.4%),10 cases were male(19.6%),36 cases were Lenke 1 and 15 cases were Lenke 2.The mean main thoracic Cobb angle and the mean thoracolumbar/lumbar Cobb angle were 44.1±7.7° and 24.1±9.3°,respectively.At the last follow-up,the wedge angle compensation of L1/2,L2/3,L3/4,L4/5 and L5/S1 segments were 3.84±5.96°,3.09±4.54°,2.30±4.53°,0.12±3.89° and-1.36±2.80° respectively,showing a decreasing trend from proximal to distal.The adjacent L1/2 and L2/3 vertebrae contributed the most to the compensation of distal unfused segments.In the represented Lenke 5&6 AIS population,37 patients developed IPCIB,and 48 patients did not.Univariate analysis showed the indexes with statistical differences between the two groups,including the number of non-fused vertebrae,preoperative main thoracic Cobb angle,preoperative thoracolumbar/lumbar Cobb angle,preoperative lumbar apical vertebral translation,preoperative coronal balance,preoperative L5 tilt angle,bending L5 tilt angle,preoperative and postoperative thoracic apical vertebral translation,postoperative thoracolumbar/lumbar Cobb angle,postoperative lumbar apical vertebral translation,postoperative radiographical shoulder height and L5 tilt angle.Logistic regression analysis showed that the main risk factors of IPCIB were the preoperative bending L5 tilt angle,postoperative thoracic apical vertebral translation,and postoperative lumbar Cobb angle.The IPCIB index was defined as 1.3*preoperative bending L5 tilt angle+1.5* postoperative thoracic apical vertebral translation-0.8*postoperative lumbar Cobb angle.When the IPCIB index was greater than 16,the incidence of IPCIB was 88% and the non-incidence was 90%.A total of 425 patients with an average age of 14.60±2.08,were included in the preliminary study on the construction of posterior surgical correction model based on deep learning network,including 77 males and 348 females.The results showed that the mean square error of the verification set was 2.7665×10-5 and the average absolute error was 0.0035 for the multi-layer depth network model;the mean square error of the verification set for the encoder-decoder model was 0.0067302433558677 and the average absolute error was 0.041656944900751114;the mean square error of the verification set for the convolution and attention mechanism model was 0.005864703182095824 and the average absolute error was 0.0397668220102787;the mean square error of the validation set for deep factorization machine model was 0.005253330513687367 and the average absolute error was 0.05162423476576805.Among them,the error analysis and curve simulation results of verification set of multilayer depth network model were better than the other three models.【Conclusions】1.After posterior selective thoracic fusion in Lenke 1&2 AIS patients,the unfused segment could achieve great spontaneous compensation.When L1 was selected as the lowest instrumented vertebra,the compensation of distal unfused lumbar segments decreased from proximal to distal,and the adjacent L1/2 and L2/3 segments made the greatest contribution to compensation,which can further provide theoretical support for the occurrence of addingon phenomenon.2.The main risk factors of IPCIB in Lenke 5&6 AIS patients were the preoperative bending L5 tilt angle,postoperative thoracic apical vertebral translation and postoperative lumbar Cobb angle.IPCIB index had the ability to accurately predict the occurrence of IPCIB.In the process of orthopedic surgery,we’d better correct the thoracic curve as much as possible,and correct the lumbar curve moderately.3.The artificial intelligence model based on deep learning can predict the coronal results after AIS posterior correction,which provided a certain reference value for the correction strategy,and it realized the simulation of the correction effect under different surgical strategies to a certain extent.In terms of the current sample size,the multi-layer depth network model had the best prediction results,and its simulation curve was better than the other three models. |
PDF Full Text Request