<p>Brain tumors are some of the most malignant neurological conditions, frequently leading to severe cognitive impairments and decreased quality of life. Timely and accurate diagnosis is important for successful treatment; however, conventional imaging-based methods suffer from shortcomings such as heterogeneity in tumor appearance, tiny differences among tumor types, and high reliance on experienced radiologists. To overcome these challenges, we obtained a large-scale MRI dataset from the Medical University of Tianjin and Nanfang Hospital in Guangzhou covering a variety of tumors and normal brain images. We introduce a new deep learning architecture incorporating an EfficientNet-B0 backbone with a Transformer module to extract both local and global contextual features. The model also includes a voting ensemble strategy that utilizes multiple architectures to enhance robustness and classification accuracy. Comprehensive assessment based on accuracy, F1-score, AUC, precision, and recall shows that the proposed solution has over 99% accuracy and is superior to the state-of-the-art. Such outcomes illustrate the ability of the framework to effectively differentiate tumor types and facilitate automated diagnostic pipelines. The research focuses on the possibility of integrating deep feature learning with Transformer-based attention mechanisms for accurate, scalable, and clinically viable brain tumor classification.</p>