The integration of Artificial Intelligence (AI) into healthcare represents a transformative sift towards smart medical systems. This paper presents an innovative framework for automated brain tumor classification by integrating multi-head attention mechanisms into pretrained Convolutional Neural Networks (CNNs) to enhance accuracy. We methodically evaluate five state-of-the-art CNN architectures e.g. VGG16, VGG19, MobileNet, Xception, and InceptionV3 enhanced with attention layers to classify MRI brain images into four categories glioma, meningioma, pituitary, and normal. To mitigate dataset imbalance, we apply RandomOverSampler and extensive data augmentation methods. Experimental findings reveal that the attention-enhanced MobileNet model attains superior performance, attaining a classification accuracy of 97%, surpassing conventional transfer learning approaches. To further interpret the model’s decisions, Explainable Artificial Intelligence (XAI) technique is employed to highlight key areas in the MRI images that influenced the classification outcomes, thereby enhancing model transparency. Overall, this work demonstrates performance improvement over other existing methods.