Vision transformer-based binary classification for automated brain tumor detection in MRI

In the clinical setting, the diagnosis and classification of brain tumors are critically dependent on accurate identification of the tumor from MRI, but manual diagnosis is time-consuming and prone to inter-observer variability. Deep learning has been extensively used in MRI analysis, but most existing Vision Transformer (ViT)-based techniques are either large and expensive or heavily rely on pretrained models, which are not optimized for small clinical datasets. To address this limitation, we conduct a systematic study of performance-efficiency trade-offs of compact Vision Transformers (ViTs) trained from scratch for binary brain tumor classification. The proposed model uses patch-based self-attention to capture global contextual relationships in MRI slices, overcoming the locality constraints of conventional CNNs. A tailored medical augmentation pipeline and a compact classification head are employed to improve generalization under limited data. Experiments on the Br35H dataset yield a test accuracy of 94.66% with an F1-score of 94.6%. The robustness of these findings is confirmed through 5-fold cross-validation (mean accuracy 94.73% ± 0.81%) and cross-dataset evaluation. To support interpretability, t-SNE visualizations and attention-rollout maps are provided, demonstrating discriminative feature separation and anatomically plausible attention focus. When trained and evaluated on the larger Brain Tumor MRI Dataset (10,093 images), the same architecture achieves 97.14% test accuracy, confirming scalability beyond the primary benchmark. The model maintains an extremely small footprint of only 0.54M parameters, 52.52M FLOPs, and approximately 2.06 MB of storage, making it suitable for deployment in resource-constrained clinical environments.