Brain tumors are unusual and uncontrollable development of brain or the central spinal canal cells that can disrupt regular brain activity and, if not identified in time, could become fatal. For diagnosis and treatment planning to be successful, tumour types must be accurately identified. The study mainly focuses on the multiclass categorisation of brain tumor Magnetic Resonance Imaging (MRI) scans into following 4 categories: pituitary tumor, meningioma, glioma, and no tumor. Using a public MRI dataset, substantial preprocessing and data augmentation techniques were applied to improve image quality and model generalization. A variety of deep learning models were trained and compared, including pretrained Convolutional Neural Networks—Residual Network-50 (ResNet50) and Visual Geometry Group-16 (VGG16)—as well as custom Shifted Window Transformer (Swin Transformer) and Vision Transformer (ViT) architectures to establish baseline performance. Additionally, ensemble strategies such as average, majority voting, weighted ensembles, and stacking with meta-learners including logistic regression, random forest, and a neural network were explored. The random forest stacking meta-learner achieved the highest accuracy of 97.41%, indicating that combining models through stacking substantially enhances classification performance over any single model, offering a robust approach for automated brain tumor diagnosis.