Construction And Application Of An Intelligent Assistant Diagnosis System For Adolescent Idiopathic Scoliosis

Objective: To construct a new interactive measurement system for adolescent idiopathic scoliosis(AIS),named "AISpine".Compared with Surgimap software,the consistency and reliability of measuring imaging parameters with AISpine system were explored.Method: 1.Based on the browser/server(B/S)architecture,an AIS interactive measurement system is constructed,which includes 29 parameters of the images of the front and side position and Bending position of the spine.2.Collect 31 patients diagnosed as AIS in our hospital from January,2018 to August,2020 through clinical and full spine X-ray films.Twenty-two parameters of the frontal and lateral X-ray films of the whole spine were measured by three surveyors using Surgimap software and AISpine system respectively.Three surveyors used two kinds of measuring tools to measure twice,and each measurement interval was 2 weeks.The measurement results and time of the three surveyors were recorded.3.The reliability of the results measured by three surveyors using AISpine system was analyzed,and the repeatability of the three surveyors’ two measurements is analyzed.The paired sample T-test or Wilcoxon signed rank sum test was used to analyze the differences between the measurement results of Surgimap software and Aisine system.Results: 1.The reliability analysis of AISpine: except SO(ICC = 0.611),the reliability of other measurement results is good or excellent.2.The repeatable analysis of Surgimap: the consistency of 19 parameters is good or excellent;the consistency of TL/L-AVT and TK is good;the consistency of SO is low;The repeatable analysis of AISpine: the consistency of 19 parameters is good or excellent;T1S and TK have good consistency;Low consistency of SO.3.Analysis of differences between groups: In the first measurement,the measurement results of Surgimap software and AISpine system used by three surveyors were compared between groups.T-AVT,TL/L-AVT and T1 SPI were significantly different(P < 0.05),but there was no significant difference in the other 19 parameters between groups(P > 0.05).4.Time analysis: The time difference between the two methods is(0.92±1.21)min and(0.82±0.96)min,respectively,with statistical significance(P < 0.05).Conclusion: Through the comparison and verification of the two measurement methods,AISpine system has good reliability and repeatability in the measurement of coronal and sagittal parameters.Especially for AIS image measurement,this system has the characteristics of high efficiency,less time consumption and user-friendly interface,which provides a new measurement tool for spinal surgeons.Objective: To construct the prediction model of Lenke classification according to the AIS classification standard put forward by Lenke in 2001,and compare the classification performance of logic regression,support vector machine and decision tree model in Lenke classification.Compared with the classification ability of graduate students in spinal surgery,the clinical application value of the optimal classification model was evaluated.Method: 1.Data preprocessing: 319 patients with AIS diagnosed by clinical and full spine Xray films from January 2013 to January 2021 in our hospital were collected.The subjects were randomly divided into training and test sets on an 8:2 scale using a computer.Two senior attending physicians used Surgimap software to measure the whole spine X-ray films of AIS patients,and recorded the Lenke classification of patients.2.Model construction: The data of model construction comes from the index data in the process of Lenke classification,including Cobb angle,main curvature,structure,and the position relationship between the vertical line of sacrum and the pedicle of lumbar vertebra.Ten subtypes of Lenke classification were selected,and logistic regression model,support vector machine model and decision tree model were trained with classification parameter data,and the best classification model was selected by test set.3.Evaluation indicators: Accuracy,precision,recall,F1 value,ROC curve and AUC value are used as indicators to evaluate the performance of different classification models.Randomly selecting the marked image data from the test set,six graduate students of spinal surgery were tested by Lenke typing.The single-sample T-test or Wilcoxon signedrank test was used to compare the classification ability of graduate students in spinal surgery with that of the optimal model.Results: 1.In the test set,the classification performance of 10 subtypes of Lenke typing.The average accuracy of logistic regression model is 0.875,and AUC value is 0.93.The average accuracy of SVM model is 0.91,and AUC value is 0.95.The average accuracy of the decision tree model is 0.95,and the AUC value is 0.97.The overall classification performance of the decision tree model is the best.2.The accuracy,recall and F1 value of Lenke classification by decision tree model are better than those of graduate students in spinal surgery,and the accuracy,recall and F1 value of graduate students,especially those in Lenke1 and Lenke2,are lower.Conclusion: Among the three machine learning classification models,the decision tree model has a good overall classification performance.Compared with the classification judgment ability of graduate students in spinal surgery,the decision tree classification model has obvious advantages,which also indicates that the model has a good application prospect in assistant diagnosis of junior doctors.The decision tree model is wrongly classified,so it is necessary to further increase the training samples and compare it with the classification ability of spinal surgeons of different levels.Objective: To explore the accuracy of automatic segmentation of AIS positive X-ray film and Cobb angle measurement based on deep learning,compare the segmentation effect of U-Net and three improved U-Net network models,and to evaluate the performance of Cobb angle measurement model by comparing with the measurement results of professional doctors.Method: 1.Data preprocessing: 163 patients with AIS diagnosed by X-ray films of the whole spine from January 2013 to January 2021 in our hospital were collected.According to the ratio of 8: 2,they are randomly divided into training set and testing set.The vertebral bodies were manually marked by two spinal surgeons using Labelme software,and the marking targets included C7-L5,18 single vertebral bodies and key points of vertebral bodies.2.Model construction: Using manually marked spinal mask map,single vertebral mask map and data sets of vertebral key points,the U-Net vertebral segmentation model,improved U-Net vertebral segmentation model and vertebral landmark detection model are respectively constructed by using training sets,and the best model is obtained by iterative training and parameter optimization.3.Evaluation index: Dice similarity coefficient,intersection ratio and accuracy rate are used to evaluate the automatic segmentation effect of different U-Net models in the test set.ICC correlation coefficient analysis professional physician and deep learning model were used to analyze the correlation of Cobb angle measurement.Use BlandAltman analysis to evaluate the consistency between measurement results.Results: 1.In terms of image segmentation: R2U-Net model is generally applicable,which can accurately identify and mark a single vertebral body,and predict that the boundary of the spliced vertebral body is smooth and complete,with the accuracy of 95.94%,DSC of 91.03% and IOU of 83.54%.Compared with the original model,DSC increased by 4.16% and IOU increased by 6.68%.2.Professional physicians and deep learning model in Cobb angle measurement.Correlation analysis of the two methods: the correlation coefficient of the measured values in PTCobb,MTCobb and TL/LCobb exceeds 0.9(P < 0.05),indicating that the two methods have significant positive correlation.Consistency analysis of the two methods: three points of Bland-Altman scatter plot basically fall within the consistency interval,which indicates that the two methods are in good consistency.The number of points where the difference of PT Cobb,MT Cobb and TL/L Cobb exceeds the 95% consistency limit is 2(7.1%),2(7.1%)and 1(3.6%),respectively.Conclusion: Among the improved U-Net segmentation models,R2U-Net model is generally applicable,which can accurately identify and segment a single vertebral body.Through image morphology processing,the boundary of vertebral body can be smooth and complete.The automatic measurement model of Cobb angle has good consistency with the manual measurement results of professional physicians,and the application of deep learning technology is expected to further improve the measurement accuracy.