Deep Hybrid Models for Multi-Class Brain Tumor Classification: Integrating CNN, Transformers, and Explainable AI for Enhanced MRI-Based Detection

Classifications of brain tumors using MRI or magnetic resonance imaging is a critical responsibility in neurooncology, directly influencing diagnosis, treatment planning, and prognosis. In like this study, We suggest a fresh deep search hybrid framework include convolutional Neural Networks (CNNs), Transformers, and Explainablinclude convolutionale AI (XAI) techniques to classify brain Low-grade glioma (LGG), high-grade glioma (HGG), tumors, and Healthy classes. Leveraging multi-contrast MRI modalities (FLAIR, T1, T1ce, T2), our approach enhances both the representational depth and clinical interpretability of tumor detection systems.We introduce and evaluate five advanced architectures: (1) YoReCNNBi-Net, a YOLO-CNN-BiLSTM hybrid that combines real-time localization with sequential learning; (2) MC-CNN-XAI, which integrates multi-contrast fusion with Grad-CAM visual explanations; (3) 3D Efficient-SE, using 3D convolution and Squeeze-and-Excitation blocks; (4) HT-CNN, a ResNet-Transformer hybrid model; and (5) RU-DenseNet, combining segmentation via Residual U-Net with DenseNet-based classification.Extensive experiments on a curated dataset of over 400 MRI volumes demonstrate that YoReCNNBi-Net outperforms all other models, achieving weighed F1-score of 0.9625 and macro F1-score of 0.9583, with class-wise F1-scores of 0.9300 (LGG), 0.9450 (HGG), and 0.9999 (Healthy). The MC-CNN-XAI model closely follows with a macro F1-score of 0.9549, showcasing the strength of multi-modal fusion and XAI-driven transparency. All models are validated using confusion matrices and ROC curves, demonstrating high sensitivity, specificity, and generalization across tumor types. By combining spatial, sequential, and contextual learning with visual interpretability, our framework sets a new benchmark in automated brain tumor classification. The proposed models not only enhance diagnostic accuracy but also offer explainable outputs, promoting trust and adoption in real-world clinical environments.