The process of classifying brain tumors through MRI scans faces difficulties because tumors have different growth patterns and there are not enough diverse datasets and researchers need to create dependable ways to represent important features. This work presents a multiclass brain tumor classification method which uses a hybrid multi-transformer feature fusion framework and Grey Wolf Optimizer (GWO) technology to control hyperparameter optimization. A unified 3584-dimensional representation is created through the extraction of deep features from four pretrained architectures which include Vision Transformer (ViT-B/16) and Swin Transformer (Swin-B) and BEiT (BEiT-B/16) and ConvNeXt (ConvNeXt-B). The GWO optimization with ResNet101 classifier enhances its ability to generalize while reducing the risk of overfitting. The work utilized a Kaggle brain MRI dataset which includes 6799 T1-weighted contrast-enhanced images that show four distinct categories of glioma meningioma pituitary tumor and no tumor. The proposed fusion-based model achieved a test accuracy of 97.73% which exceeded the performance of both individual backbone features and baseline classifiers. The ablation analysis results combined with the statistical evaluation demonstrate that feature fusion and metaheuristic optimization lead to performance improvement. The results show strong performance on the tested dataset, but clinical application requires validation through external testing at multiple centers in future.
Saravanan et al. (Sat,) studied this question.