Prediction Of MGMT Status In Glioblastoma Via Deep Learning And Radiomics

[Objective]The purpose of this study is to explore the value of radiomics and deep learning methods in the prediction of methylation status of the O~6-methylguanine methyltransferase(MGMT)in glioblastoma,and to build an automatic,convenient,low-cost and non-invasive prediction model,so as to contribute to the personalized treatment of glioblastoma.[Materials and methods]The methylation status of the MGMT gene promoter and the preoperative MRI data of 106 patients with glioblastoma were retrospectively collected.Finally,87CE-T1WI and 66 FLAIR cases were included.On the one hand,the region of interest(ROI)was obtained by manually segmented the maximum boundary of the tumor slice by slice.After preprocessing,radiomics features were extracted from MRI.Then,the features selection were by principal component analysis(PCA)and analysis of variance(ANOVA).A support vector machine(SVM)was used to construct a prediction model for the methylation state of MGMT gene promoter.Finally,The accuracy,sensitivity,specificity,positive predictive value and negative predictive value of each model and the maximum area under the curve(AUC)were evaluated by drawing the receiver operating curve(ROC).On the other hand,The end-to-end pipeline completes both tumor segmentation and status classification.In the automatic segmentation module,we added a variational autoencoder(VAE)branch to the full convolutional network(FCN)model,and then calculating the Dice coefficient to evaluate the performance of tumor segmentation.Meanwhile,for the evaluation of time consumption,we recorded the total time and divided it by the number of slices.For the MGMT methylation status classification,we designed a4-layer convolutional neural network(4-CNN)model by ourselves,and then cascaded the model with the automatic segmentation module.the performance of the prediction model was evaluated by calculating the accuracy,recall rate,precision,F1 score and maximum area under the curve(AUC)of the model in the training set and the validation set.[Results]1.The methylation status of MGMT promoter in glioblastoma can be predicted by using traditional radiomics features.The prediction level of traditional radiomics model based on manual segmentation on CE-T1WI and FLAIR images is close(the Accuracy is 0.6898,0.6861;the AUC is 0.6593,0.6861).2.The automatic segmentation model segmented by conventional MRI deep learning method has better tumor segmentation performance.Compared with CE-T1WI(Dice score,0.828),the model on FLAIR(Dice score,0.897)showed better performance.Compared with manual segmentation,automatic segmentation model has a faster segmentation speed(the average automatic segmentation time of one slice on CE-T1WI and FLAIR sequence is 0.11 s and 0.07 s,corresponding to the average manual segmentation time of 50 s and 60 s).3.Conventional MRI deep learning method can be used to predict the methylation status of MGMT promoter in glioblastoma.Compared with CE-T1WI,the deep learning model based on FLAIR have better tumor prediction performance(the accuracy is 0.804,0.827;the AUC is 0.8866,0.9052).[Conclusion]Both deep learning and radiomics are of great value in predicting the methylation status of MGMT promoter in glioblastoma,but the predictive value of deep learning method is better.Compared with traditional radiomics based on manual segmentation,deep learning has more advantages for automatic tumor segmentation.It can save time,avoid inter-observer differences in tumor segmentation,and facilitate the discovery of molecular biomarkers from conventional medical images to facilitate diagnosis and treatment decision-making.