The accurate and timely identification of brain tumors is crucial for effective diagnosis and treatment planning; however, the manual interpretation of MRI scans continues to be difficult and susceptible to errors. Although convolutional neural networks (CNNs) have made strides in automated classification, their dependence on local feature processing can restrict overall effectiveness. As an initial exploration, this pilot study introduces a Vision Transformer (ViT) model that utilizes self-attention mechanisms to capture both long-range global contexts and detailed local dependencies within image data, facilitating a more thorough feature representation that is vital for detecting subtle pathological patterns. Trained and assessed on a pilot dataset comprising 3,000 MRI images with significant augmentation, the proposed ViT model attained a promising preliminary accuracy of 99.73%, surpassing established CNN-based architectures such as ResNet-50, VGG-16, and EfficientNet-B0 across all evaluation metrics within the constraints of this binary classification task. These feasibility results not only highlight the potential of ViTs for brain tumor classification but also effectively validate the fundamental data processing and model fine-tuning pipeline. The study points out critical limitations, including dataset scale and model explainability, which directly influence the design of a forthcoming large-scale, multi-institutional research initiative. This pilot research lays a foundational framework for the integration of transformer-based models into medical imaging workflows to enhance diagnostic accuracy.
Asuai et al. (Sun,) studied this question.