<p>A brain tumor is an abnormal growth of cells in the brain, critical for diagnosis and treatment. The rising incidence of brain tumors highlights the need for early identification, as their diversity and complexity make classification challenging. Early detection and accurate classification systems are essential for improving treatment outcomes. Several existing studies aimed to improve diagnostic accuracy by combining predictions or features from various CNN models, but these methods often caused discrepancies and inconsistent results. Utilizing features from various models can significantly enhance the stability and accuracy of tumor classification, but crucial features are often overlooked. To address this issue, our study presents an efficient framework for brain tumor classification that integrates feature level fusion with a novel self-attention mechanism for optimal feature refinement. We leverage lightweight pre-trained architectures named MobileNetV1 and MobileNetV2 due to their exceptional performance and efficient design of using depthwise separable convolutions. These models serve as the base models of our proposed framework, and feature fusion is performed by effectively combining their unique features to enhance performance. To further strengthen the focus of proposed model on key features, we developed a novel lightweight self-attention mechanism that selectively prioritizes the most important features, thus improving diagnostic accuracy. To evaluate our approach, we utilized a publicly available Kaggle brain tumor dataset that includes four distinct classes: glioma, meningioma, no tumor, and pituitary. Our proposed approach not only achieved high accuracy of 98.55% but also maintained the smallest model size of 38.50 MB among competitors. Additionally, we employed Grad-CAM and feature map Visual analysis to highlight critical areas and show how the model layers process and refine diagnostic features.</p>