Surgical Outcome Prediction For Chiari Ⅰ Malformation Syringomyelia Based On T2[*] Weighted Radiomics Features

Objective(s): This study aimed to construct a machine learning-based T2[*]-weighted image(WI)radiomics model for predicting postoperative recovery in patients with CM-I syringomyelia.Methods: 125 patients with CM-Ⅰ syringomyelia underwent preoperative T1[*]and T2[*] weighted images(WI),and all underwent posterior fossa decompression(Posterior Fossa Decompression,PFD).They were divided into good/poor outcome groups based on recovery rate.The whole dataset was randomly divided into training set(N=88)and test set(N=37),and then in T2[*]WI,the syringomyelia site was segmented layer by layer as a region of interest(ROI)and the radiation group was extracted academic features.The random forest algorithm(Random Forest,RF)and logistic regression(Least absolute shrinkage and selection operator,LASSO)were used to screen the feature variables to obtain the best radiomics features.Morphological parameters of the posterior fossa and linearity of the syringomyelia were selected as routine MRI features on T1[*]WI.Clinical characteristics were age,total disease duration,length of hospital stay,and preoperative pain and nonpain symptoms.Four models were constructed for performance evaluation: radiomic,radiomic,clinical-radiomic,and clinicalradiologic.Five-fold cross-validation was used for model evaluation and calibration,and the area under the receiver operating characteristic curve(Area under the ROC curve,AUC)was used to evaluate the performance of the model.Based on the De Long test,an AUC of P<0.05 was considered a meaningful performance predict.Results: 1.Radiology model AUC(training set: 0.859,test set: 0.426),accuracy0.553,sensitivity 0.333,specificity 0.696;radiomics model AUC(training set: 0.901,test set: 0.817),accuracy 0.816,sensitivity 0.733,specificity 0.869;clinicalradiomics model: AUC(training set: 0.871,test set: 0.699),accuracy 0.594,sensitivity 0.673,specificity 0.704.Clinical-radiology model: AUC(training set:0.861,test set: 0.461),accuracy 0.372,sensitivity 0.365,specificity 0.653.The radiomics model showed good predictive ability,followed by the clinical-radiomics model,and the remaining two models showed poor predictive performance.2.The 62 features of the radiomics construction model are divided into: 10 shape features,13 first-order features,11 gray-scale co-occurrence matrix features,7 grayscale area size matrix features,and 12 gray-scale features according to their unique characteristics.There are 2 features of degree stroke matrix,2 features of neighborhood gray level difference matrix and 7 features of gray level correlation matrix.3.The top three important features of the radiomics model are: shape featuremaximum axis length,first-order feature-robust mean absolute deviation and gray level correlation matrix-large dependence and high gray level emphasis;clinicalradiomics model The top three important features are: shape feature-maximum axis length,clinical feature-total disease duration(months)and gray travel moment positive feature-gray level inhomogeneity.The shape feature-maximum axis length is the most important feature in both radiomics and clinical-radiomics.Conclusion(s): The radiomics features derived from MRI raw images can be used as predictors of postoperative recovery in CM-I syringomyelia;among them,the shape feature-maximum axial length is the most important feature of radiomics.Radiomics may have broader advantages over traditional MRI metrics.