| Objective: In this study,a comprehensive machine learning（ML）model was developed to predict the isocitrate dehydrogenase（IDH）mutation based on the preoperative conventional magnetic resonance imaging（MRI）of lower-grade gliomas（LGGs）.Materials and methods: 1.Retrospectively collected 227 patients with LGGs confirmed by pathological results in Lanzhou University Second Hospital,and divided the patients into training cohort（n=160）and validation cohort（n=67）in a ratio of 7:3.The preoperative visually accessible Rembrandt images（VASARI）features of patients were extracted,and the clinical and VASARI features associated with IDH mutation were selected by univariate analysis,using the least absolute shrinkage and selection operator（LASSO）regression analysis screened out VASARI features closely related to IDH mutations;based on clinical and VASRI features,a clinical-VASARI model for predicting IDH mutations in LGGs was developed.Receiver operating characteristic curve（ROC）and area under the curve（AUC）values were used to evaluate the predictive performance of the model.2.Retrospectively collected 136 patients with LGGs confirmed by pathological results in the Lanzhou University Second Hospital.All patients had complete and high-quality preoperative routine MRI,including T1-weighted imaging（T1WI）and T2-weighted imaging（T2WI）,T1 weighted contrast-enhanced imaging（T1C）and T2 fluid attenuated inversion recovery sequence（FLAIR）.Enrolled cases were divided into training cohort（n=96）and validation cohort（n=40）in a ratio of approximately 7:3.Radiomics features were extracted from the above four MR preoperative sequences,and the Radiomics features closely related to IDH mutation in each sequence were screened by univariate analysis and LASSO method as radiomics signatures.A multi-sequence fusion-based radiomics fusion model（All-model）was constructed using the multi-factor stepwise logistic regression method.Univariate and multivariate analysis were used to screen VASARI features associated with IDH mutations,and a clinical-radiomics comprehensive model（COMB-model）was developed by integrating VASARI features and Radiomics fusion model,and on this basis,a nomogram was constructed.The predictive performance and clinical utility of each model were evaluated using ROC curves and AUC values,calibration curves（CC）and decision curves（DC）.3.A total of 552 MR scan sequences of 138 patients with complete preoperative conventional MRI（T1WI,T2 WI,T1C and FLAIR）were retrospectively collected from the Lanzhou University second hospital.The proportions were randomly divided into training set（n=115）and validation set（n=23）for six-fold cross-validation to train the model.Based on the above four sequences of MRI,MR single-sequence and multi-sequence three-dimensional densely connected convolutional network（3D-Dense Net）structure deep learning（DL）prediction models were developed respectively.Applying ROC curve and AUC value,accuracy,sensitivity,specificity,positive predictive value（PPV）,negative predictive value（NPV）,false positive rate（FPR）,false negative rate（FNR）,false discovery rate（FDR）,F1 value,confusion matrix to evaluate the predictive performance of each model and compare them.Results: 1.Among the preoperative VASARI features,11 VASARI features were correlated with IDH mutation status after univariate analysis.These 11 VASARI features were reduced by LASSO regression to retain 5 features closely related to IDH mutation,their names and coefficients were F1（Tumor Location）,-0.315;F6（Proportion n CET）,0.171;F11（Thickness of enhancing margin）,-0.168;F20（Cortical involvement）,-0.037 and F30（Lesion Size Y）,0.004.Based on these five VASARI features,we constructed a prediction model for LGGs IDH.The AUC,sensitivity and specificity of the model in the training set were 0.801,0.840 and 0.722,respectively;the AUC,sensitivity and specificity of the prediction model in the validation set were 0.752,0.921 and 0.517,respectively.2.From each A total of 4148 radiomic features were extracted from the four MRI sequences of the patient,and 11 features significantly related to LGGs IDH mutation were further screened（T1WI: 2,T1C: 3,T2WI: 4,FLAIR: 2）.Among the single-sequence model and the All-model,the latter had the best performance,with AUC,accuracy,and sensitivity of 0.896,0.802,0.710 and 0.843,0.725,0.615 in the training and validation sets,respectively.However,the COMB-model developed by combining VASARI features and radiomics features had the best predictive performance,with AUC values,accuracy and sensitivity of 0.925,0.865,0.806 and 0.907,0.800,0.769 in the training set and validation set,respectively.Calibration curve analysis showed that the IDH mutation probability predicted by COMB-model was in good agreement with the actual mutation probability.Clinical decision curve analysis also showed that the COMB-model had the best net benefit.3.Among the 3D-Dense Net models constructed based on T1 WI,T2WI,T1 C and FLAIR,the prediction performance of each sequence model was different.Among them,the 3D-Dense Net model developed with the FLAIR sequence performed the best.The FLAIR-based 3D-Dense Net model achieved a prediction accuracy of 0.800（AUC = 0.953）in the training set and 0.703（AUC = 0.773）in the validation set.The best accuracy of the multi-sequence fusion model was 0.922（AUC = 0.998）in the training set and 0.783（AUC = 0.854）in the validation set.The prediction performance of the multi-sequence fusion model is better than that of the single-sequence model.Conclusions: 1.The VASARI features,radiomics and deep learning model based on conventional MRI can non-invasively predict the IDH mutation status of LGGs.Of all the models,the deep learning model had the best predictive performance.2.Among the machine learning models,the multi-parameter and multi-sequence fusion model has better predictive performance and can be used as an important auxiliary means for non-invasive preoperative prediction of LGGs IDH mutation status. |
