Grading diffuse glioma based on 2021 WHO grade using self-attention-base deep learning architecture: variable Vision Transformer (vViT)

To evaluate the diagnostic performance of the self-attention-based model, termed variable Vision Transformer (vViT), in the task of predicting the grade of diffuse glioma based on the 2021 World Health Organization (WHO) central nervous system (CNS) tumor classification. This cross-sectional study analyzed adult patients with histopathologically confirmed diffuse glioma, following the 2021 WHO CNS tumor classification. We used age, sex, radiomic features, and four MRI sequences to predict the grade of gliomas. As binary classifications, we constructed three models: 2 vs. 3/4 (326 patients with 1575 grade 2 and 1574 grade 3/4 images), 3 vs. 2/4 (330 patients with 1726 grade 3 and 1726 grade 2/4 images), and 4 vs. 2/3 (333 patients with 3292 grade 4 and 3292 grade 2/3 images). As a multiclass classification, we constructed a 2 vs. 3 vs. 4 model (334 patients with 1575 grade 2, 1575 grade 3, and 1575 grade 4 images). Metrics including accuracy and area under the curve of the receiver operating characteristic (AUC-ROC) were calculated. The highest accuracy and AUC-ROC were 0.84 (95% confidence interval; 0.75–0.93) in multiclass classification (2 vs. 3 vs. 4) and 0.94 (0.88–0.98) in 4 vs. 2/3, respectively. The highest Cohen’s κ coefficient between ground truth and the predicted value was 0.54 obtained in the multiclass classification (2 vs. 3 vs. 4). The vViT is a competent multi-modal deep-learning model that can predict the grade of gliomas which were classified based on the 2021 WHO CNS tumor classification.