Comparative Evaluation of Transformer-Based Architectures for Brain Tumor MRI Classification: Balancing Predictive Performance and Computational Efficiency

Abstract

Reliable categorization of brain tumors on magnetic resonance imaging (MRI) plays an important role in radiological decision support for differentiating tumor subtypes with overlapping visual characteristics. This study investigates a transformer-based transfer learning framework for classifying brain MRI scans into four clinically relevant categories: glioma, meningioma, no tumor, and pituitary tumor. The experimental dataset comprised 5,941 brain MRI scans, which were partitioned into training, validation, and independent test subsets at an approximately 70:15:15 ratio, yielding 4,160, 892, and 889 images, respectively. A consistent experimental protocol was applied to evaluate four transformer architectures: Swin-Base, DeiT-Base, ViT-Base, and BEiT-Base. The evaluation combined predictive and computational criteria, including accuracy, macro-level precision, recall, F1-score, AUROC, parameter count, GFLOPs, per-image inference time, and image throughput. Among the evaluated models, Swin-Base achieved the highest overall classification performance, with an accuracy of 0.9854, macro-averaged precision of 0.9832, macro-averaged recall of 0.9855, and macro-averaged F1-score of 0.9843. BEiT-Base achieved the highest macro-averaged AUROC value of 0.9980, indicating strong class separability across decision thresholds. The confusion matrix of Swin-Base further demonstrated robust class-wise discrimination on the 889-image independent test set, correctly classifying 876 images. These findings suggest that transformer-based architectures, particularly Swin-Base, may provide a highly effective approach for automated brain MRI-based tumor classification.