Brain tumor classification using fractional Laplacian image enhancement and vision transformer

Early diagnosis and successful treatment require accurate identification of brain tumors using magnetic resonance imaging (MRI). However, MRI slices often exhibit low contrast, blurred boundaries, and noise, which can negatively affect automated classification performance. Traditional improvement processes can result in excess sharpening of edges or altered structural continuity, which could affect subsequent feature extraction. In this regard, a hybrid method combining fractional Laplacian-based image enhancement with vision transformer (ViT) is developed in two stages. The fractional Laplacian processing enhances the boundaries between structures and preserves structural continuity in the first stage. The second stage is the vision transformer (ViT), which captures local and global dependencies in order to do multi-class classification effectively. The structural similarity (SSIM) and the entropy are used to assess the frameworks by enhancement metrics, and the Accuracy, Precision, Recall, F1-score, and ROC-AUC are used to determine the classification performance. Experiments are carried out using more than one random seed and reported as mean as standard deviation and statistically tested. The highest configuration had a mean test accuracy of $$98.46 \% \pm 0.29$$ and the statistical test showed no significant difference between raw and enhanced inputs ($$p > 0.05$$). Competitive classification performance is shown to be achieved with experimental results under controlled multi-seed evaluation, and therefore, systematic analysis of the influence of fractional improvement on transformer-based classification models.