Comparative analysis of transformer architectures for brain tumor classification
Yiğitcan Çakmak, İshak Paçal
Abstract
Aim: Early and accurate diagnosis of brain tumors is critical for treatment success, but manual magnetic resonance imaging (MRI) interpretation has limitations. This study evaluates state-of-the-art Transformer-based architectures to enhance diagnostic efficiency and robustness for this task, aiming to identify models that balance high accuracy with computational feasibility. Methods: We systematically compared the performance and computational cost of eleven models from the Vision Transformer (ViT), Data-efficient Image Transformer (DeiT), and Hierarchical Vision Transformer using Shifted Windows (Swin) Transformer families. A publicly available four-class MRI dataset (Glioma, Meningioma, Pituitary, No Tumor) was used for multi-class classification. The dataset was partitioned using stratified sampling and extensively augmented to improve model generalization. Results: All evaluated models demonstrated high accuracy (> 98.8%). The Swin-Small and Swin-Large models achieved the highest accuracy of 99.37%. Remarkably, Swin-Small delivered this top-tier performance at a fraction of the computational cost of the Swin-Large model, which is nearly four times its size and with more than double the inference speed (0.54 ms vs. 1.29 ms), showcasing superior operational efficiency. Conclusions: The largest model does not inherently guarantee the best performance. Architecturally efficient, mid-sized models like Swin-Small provide an optimal trade-off between diagnostic accuracy and practical clinical applicability. This finding highlights a key direction for developing feasible and effective AI-based diagnostic systems in neuroradiology.