Evaluation Of MGMT Promoter Methylation Status In Glioblastoma By Deep Learning Based On Multimodal MRI

Objective:1.Preoperative evaluation of glioblastoma(glioblastoma)by exploring magnetic resonance imaging(MRI)image features Application value of O6-methylguanine-DNA methyhransferase(MGMT)promoter methylation status in patients with GBM.2.The deep learning model EfficientNet-B7 is Efficientnet-B7 to predict the methylation status of MGMT promotors in patients with GBM based on preoperative multimodal MRI,and the predictive performance of different modal integration methods is compared.Methods:1.Clinical and imaging data of 83 patients with GBM confirmed by surgery and pathology were collected,including 43 patients with MGMT promoter methylation and 40 patients with non-methylation.All patients underwent preoperative MRI scan.The scanning sequences included T1 weighted imaging(T1WI),T2 weighted imaging(T1 weighted imaging,T1WI),T1WI),T2-fluid attenuated inversion recovery imaging(T2-FLAIR),diffusion-weighted imaging,DWI)and three dimensional T1 contrast enhancement weighted imaging(3D-T1CE).2.By analyzing the apparent diffusion coefficient(ADC)images obtained by two imaging doctors and post-processing of DWI sequences,The location of tumor onset,characteristics of each sequence signal,ADCmin value and ADCmean value,qualitative and quantitative indicators of MR Signs were interpreted and statistically analyzed,and the differences in MR Signs between MGMT promoter methylation and non-methylation groups were compared.In addition,two imaging physicians determined the methylation status of MGMT promoter through MRI sections without knowing the pathological type of tumor.The diagnostic efficiency of radiologists was evaluated by receiver operating characteristic(ROC)Curve,Area Under ROC Curve(AUC),sensitivity,accuracy and specificity.2.Multimodal MR Images of 129 patients with GBM confirmed by surgery and pathology were retrospectively analyzed,including 3D-T1CE sequences,T2-FLAIR sequences and ADC images.Based on EfficientNet-B7 The convolutional neural network constructed T1C-net based on 3D-T1CE sequence,FLAIR-net based on T2-FLAIR sequence,ADC-NET based on ADC image,Ts-net based on three kinds of sequence superposition,IF-net based on three kinds of sequence image-level fusion,and DF-net based on three kinds of sequence decision level fusion respectively.In accordance with 7:The 129 patients were randomly divided into a training set(n=90)and a validation set(n=39).Imaging physicians manually selected all the foci level image input networks for each patient,and the predictive efficacy of the model was evaluated by ROC curve,AUC value,sensitivity,accuracy and specificity.The class activation diagram is used to visualize the focus area of the deep learning model and analyze the differences between different models.Results:1.The ADCmin and ADCmean values of tumor entity were significantly different between the two groups(P<0.05).There was statistical difference in age between the two groups(P<0.05).There were no significant differences in gender,tumor location,sequence signal characteristics,qualitative and quantitative indicators of MR Signs between the two groups(P>0.05).The accuracy of two imaging doctors was 65.1%and 69.8%,respectively.The AUC values were 0.637 and 0.684,respectively.2.The AUC values of the T1C-net,Flflair-net and ADC-net models on the verification set were 0.749,0.754 and 0.685,respectively;The AUCs of Ts-net,IF-net and DF-net models in the verification set were 0.797,0.820 and 0.783,respectively.The image-level fusion model has the highest efficacy of IF-net,while the ADC image sequence training model has the lowest efficacy of ADC-Net.Conclusions:1.The ADCmin and ADCmean values of tumor entity have certain reference value for distinguishing MGMT promoter methylation;2.The patients with MGMT promoter methylation were older,and the difference was statistically significant.Conventional MRI signs have no reference value in the diagnosis of MGMT promoter methylation status in glioblastoma.The diagnostic efficacy of imaging physicians in evaluating the MGMT promoter methylation status in glioblastoma is low.2.The deep learning model based on the multi-mode MRI EfficientNet-B7 neural network shows high diagnostic efficiency in predicting the MGMT promoter methylation status in glioblastoma.Among them,IF-net based on three sequential image-level fusion models has the best prediction performance.