Modern Vision Transformer Models for Accurate and Explainable Brain Tumor MRI Classification

Abstract

Accurate brain tumor classification from magnetic resonance imaging (MRI) remains difficult because glioma, meningioma, and pituitary tumors may exhibit overlapping enhancement, morphology, and slice-dependent appearance. This study presents a controlled comparison of Vision Transformer, Data-efficient Image Transformer (DeiT), Swin Transformer, BEiT, and EVA-02 under a common image-level protocol. The public Epic and CSCR Hospital Dataset record describes 12,064 preprocessed T1-weighted contrast-enhanced MRI images and was evaluated according to its published partitioning scheme. The original test set was reserved for final evaluation, and a validation subset was drawn solely from the original training set. BEiT-Base produced the highest observed accuracy (0.9921; Wilson 95% confidence interval, 0.9877-0.9950), a macro F1-score of 0.9924, and a macro area under the receiver operating characteristic curve of 0.9995, at a profiled cost of 33.70 giga floating-point operations per image, equal to that of Vision Transformer-Base and DeiT-Base. Accuracy intervals overlapped across models, and the differences were interpreted descriptively. Gradient-weighted Class Activation Mapping showed activation near the visually apparent lesion in many cases, but peripheral and non-lesion responses were also observed. Because patient and examination identifiers are unavailable, the findings represent image-level benchmark performance and do not establish patient-level generalization.