Tumors of the brain constitute a great importance medically as they are associated with high rates of mortality and much neurological damage. The chances of timely diagnosis are often betrayed by the vague or non-typical readouts of early signs and symptoms of tumorous effects. This work proposes a deep learning-based hybrid system for the accurate detection and classification of brain tumors from the MRI images. This research uses Vision Transformer (ViT) and GRU along with a GoogleNet-SVM hybrid model to characterize cancers into four types: no tumor, glioma, meningioma, and pituitary tumor by utilizing both spatial and temporal information for the classification. Key processes involved are data preprocessing (resizing and augmentation, and normalization), advanced feature extraction based on using the pretrained models, and temporal attention modeling. The ViT–GRU showed relatively good results, while the GoogleNet-SVM model used PCA for dimensionality reduction enabling more efficient classification. The models were trained and tested using an MRI dataset constructed for this purpose. In addition, accuracy, precision, recall, and F1-measure were calculated. t-SNE and attention heat maps are some of the techniques used on the MRI scans to demonstrate features and the effectiveness of classification. The results show that the system proves its usefulness in classification of brain tumors and hence can be useful in automated diagnosis methods in a non-invasive reliable manner. This research sets the foundation for the application of deep learning into real-world processes in the clinical field for faster and better quality pre detection and care.