Enhancing Brain Tumor Classification Using Data-Efficient Image Transformer

Medical image analysis is essential in the diagnosis and treatment of many diseases, including brain tumors. This work investigates the effectiveness of transformer models in the multiclass classification of brain tumor MRI images. We use a combination of three publicly available brain tumor datasets classified into four categories and appropriately modify existing transformer variants, including adjustments to patch size, number of categories, dropout, etc. The results indicate that, due to the condition of insufficient data volume, most transformer variants are still inferior to the benchmark model VGG-16. The DeiT model with knowledge distillation has significantly improved various indices on the test set, verifying its effectiveness in the classification of brain tumor MRI images under limited data. The article supports the application of transformer-based models to medical image analysis. Follow-up work will verify the generalization performance of the model on other datasets and explore methods to reduce the dependence of the transformer on the data, making it more applicable to tasks such as medical images classification.