Amide Proton Transfer-Weighted Imaging Of Glioma Subtypes: A Retrospective Study On Adult Patients With Diffuse Gliomas

Objective:To investigate the value of amide proton transfer-weighted（APTw）assessing the tumor region（TR）and peritumoral region（PTR）of gliomas in predicting the diffuse gliomas subtype.Establish a deep learning network model based on APTw images to help predict tumor molecular subtypes non-invasively before surgery.Methods:This retrospective study included 78 patients with diffuse gliomas(29Oligodendrogliomas,IDH^mut/1p19q^codel,30 Astrocytoma,IDH^mut,and 19 Glioblastoma,IDH^wt).Part of the study compared the difference of APTw5,APTw50（APTw mean）,APTw90,and ADC5,ADC50（ADC mean）,ADC90 values of different subtype of gliomas by histogram analysis.Multiclass receiver operating characteristic（ROC）curve analysis was used to analyze different quantitative parameters of APTw and ADC imaging.Multiclass logistic regression models were established to predict glioma subtypes.Another part of the study,we extracted the slices with tumor region,split them into training set（n=214）,test set（n=71）,and validation set（n=65）in the ratio of 60%:20%:20%.After preprocessing the MRI images,a 34-layer residual network（Res Net34）was applied to classify adult diffuse glioma.Compare three models classification accuracy of the different input methods:model 1 input nulling matrix,T1c and ADC image;model 2 input APTw image,T1-weighted post-contrast（T1c）image and ADC image;model 3 input APTw image,T1c,ADC image and clinical information,imaging morphological features and quantitative data.ROC curve and 95%confidence interval were used to evaluate the model.Grad-CAM heatmap for deep learning model visualization.Values of p<0.05 were considered statistically significant.Result:APTw90 can well distinguish three subtypes of glioma（p<0.05）.The model,which combining ADC,APT,morphological features and clinical information,fared better when using the multi-classification ROC curve to assess the accuracy of various parameters（accuracy=0.83,95%CI:0.64-0.92）,with an AUC of 0.95（95%CI:0.88-0.98）.The overall accuracy（72.31%）of the Res Net34 model combined with APTw,ADC,T1 enhancement,morphological features and quantitative data was significantly improved compared with other models.Conclusions:APTw imaging is better suited as a trustworthy biomarker for distinguishing glioma molecular subtypes as compared to ADC.A multimodal Res Net34 deep learning model based on APTw images provides another reliable option for preoperative noninvasive typing.