This research features employing attention-augmented MobileNetV2 architecture for classification of brain tumors from MRI scans, particularly glioma, meningiomas, and pituitary tumors. The model incorporates attention mechanism and uses saliency map to improve feature extraction and interpretability by allowing clinicians to visualize the important regions along which their predictions are generated. Stratified k-fold cross-validation was applied along with data augmentation for the purpose of establishing a well-performing model, while a cosine annealing learning rate schedule ensured convergence while preventing overfitting. The metrics including ROC AUC, precision, recall, Cohen’s Kappa and MCC quantify the model’s effective classification performance, collectively culminating in a ROC-AUC of 0.9703. Future research will focus on consolidating multi-modal data, which will add to diagnostic accuracy and bring about federated learning to solve privacy concerns and enhance generalization, supporting its use as a clinical decision-making tool.