Accurate and rapid detection of brain tumors in MRI images is one of the major challenges in the field of medical imaging and clinical decision-making. The limitation of labeled data and the structural diversity of tumors highlight the necessity of efficient, interpretable, and robust deep learning models that can handle data scarcity. In this study, a hybrid deep learning model for brain tumor classification is proposed, which integrates the architectures of ResNet50 and DenseNet121 along with the CBAM attention module. The ResNet50 architecture, through residual connections, enables deep and stable learning, while DenseNet121, with its densely connected structure, facilitates more effective feature transfer and improves the overall performance of the model. The CBAM module, by applying channel and spatial attention, directs the model’s focus to critical image regions, thereby enhancing classification accuracy.In the proposed model design, strategies have been employed to strengthen the model’s robustness against limited data, improving its stability and generalizability. A review of previous studies indicates that the integrated use of these three components can significantly boost the performance of neural networks in brain tumor MRI classification. Furthermore, this research addresses the challenges arising from data limitations and explores solutions such as transfer learning, data augmentation, and the incorporation of uncertainty estimation.