Fig. 4: The DL workflow for classifying molecular subtypes and grading of adult-type diffuse gliomas.

a Workflow for slice selection from multiparametric MRI (T1/T1C/T2/FLAIR) using nnU-Net to predict tumor regions. The top n% of slices by tumor area are selected, with a single slice randomly chosen per patient for model training. Traditional deep learning models based on b CNN and c visual transformer. d Overview of GlioMT for classification of adult-type diffuse gliomas according to the 2021 WHO classification. MR images are encoded using a visual transformer, while clinical data are processed by LLM-based text encoder, specifically BERT. During training, all layers of the BERT model are frozen.