A multitask classification framework based on vision transformer for predicting molecular expressions of glioma.

PURPOSE: The purpose of this study is to develop a Vision Transformer model with multitask classification framework that is appropriate for predicting four molecular expressions of glioma simultaneously based on MR imaging. MATERIALS AND METHODS: A total of 188 glioma (grades II-IV) patients with an immunohistochemical diagnosis of IDH, MGMT, Ki67 and P53 expression were enrolled in our study. A Vision Transformer (ViT) model, including three independent networks based on T2WI, T1CWI and T2 + T1CWI (T2-net, T1C-net and TU-net), was developed for the prediction of four glioma molecular expressions simultaneously. To evaluate the model performance, the accuracy rate, recall, precision, F1-score, and area under the receiver operating characteristic curve (AUC) were calculated. RESULTS: The proposed ViT model achieved high accuracy in predicting IDH, MGMT, Ki67 and P53 expression in gliomas. Among the three networks using the ViT model, TU-net achieved the best results with the highest values of accuracy (range, 0.937-0.969), precision (range, 0.949-0.972), recall (range, 0.873-0.991), F1-score (range, 0.910-0.981) and AUC (range, 0.976-0.984). Comparisons were also made between our ViT model and convolutional neural network (CNN)-based models, and the proposed ViT model outperformed the existing CNN-based models. CONCLUSION: Vision Transformer is a reliable approach for the prediction of glioma molecular biomarkers and can be a viable alternative to CNNs.