Magnetic Resonance Imaging (MRI) is an important diagnostic and treatment technique in brain tumor diagnosis and management because MRI can provide detailed images of brain tissues. Although manual interpretation of MRI scans by radiologists is time-consuming and variable, automated approaches are needed to increase precision. To tackle the problem of the accurate detection and segmentation of brain tumors, this study introduces an advanced deep learning framework based on Convolutional Neural Networks (CNNs), U-Net architectures and hybrid CNN-LSTM models. The proposed model is based on leveraging the BRATS dataset for multi-modal imaging, with robust preprocessing, data augmentation and transfer learning to solve challenges of small data, heterogeneity in tumor features. Experimental results show that the proposed models outperform traditional models in terms of detection metrics such as a recall of 94.3% and AUC-ROC of 0.96. Grad-CAM visualizations are integrated to enable interpretability, providing confidence in clinical applications, as a result. We present a highly scalable tool that improves the reliability and efficiency of MRI-based brain tumor diagnosis and aids radiologists in real-time clinical settings.