Being able to detect a brain tumor as early as possible is an important factor for the treatment plan and, thus, the survival of the patient. Brain tumor detection using deep learning techniques has significant potential in enhancing diagnostic processes through automated MRI analysis. This paper explores several state-of-the-art deep learning architectures on the BraTS 2019 dataset, including Convolutional Neural Networks (CNN), VGG16, and InceptionResNetV2. We also introduce a hybrid model that combines InceptionResNetV2 with a Transformer architecture to improve classification performance. A key aspect of our approach involves feature extraction using a hybrid model that integrates CNN with multi-head attention from Transformers, enabling effective capture of both local and global feature representations. We apply robust pre-processing techniques and evaluate the models using metrics such as F1 score, confusion matrix and sensitivity. Experimental results demonstrate that the hybrid InceptionResNetV2-Transformer model significantly outperforms traditional models, achieving superior accuracy and robustness. Our findings suggest that this hybrid model offers meaningful contributions to the advancement of brain tumor classification and diagnosis.