Abstract: Aiming at the problems of incomplete deep network feature extraction and inherent uncertainty of the model in the current deep learning model for the prediction of glioma isocitrate dehydrogenase 1(IDH1)gene status based on Magnetic Resonance Imaging(MRI), the truth discovery divided attention-ResNet(TDA-ResNet)model are proposed based on improved residual network(ResNet)with truth discovery. Firstly, the ResNet network model architecture is optimized by the divided attention mechanism to extract local and global features of glioma images to predict the IDH1 gene status of glioma; meanwhile, the truth discovery algorithm is incorporated into the model to calibrate the uncertainty of the depth feature vectors as the prediction results, so as to improve the prediction accuracy of the model. The experimental data were collected from the MR images of some glioma patients in the Affiliated Hospital of Xuzhou Medical University and the public dataset of the cancer imaging archive(TCIA). The experimental accuracies of the TDA-ResNet model in the MR image dataset of glioma in the Affiliated Hospital of Xuzhou Medical University and the TCIA dataset were 95.73% and 94.3%. The experimental results show that the TDA-ResNet model can achieve non-invasive prediction and uncertainty calibration of IDH1 gene status of glioma, and its performance is better than the existing deep learning prediction model of IDH1 gene status, which is of great significance for the clinical diagnosis and treatment of glioma.

Key words: glioma, isocitrate dehydrogenase 1, deep learning, truth discovery, uncertainty calibration