Precise MRI-related brain tumor segmentation is critical to efficient diagnosis planning as well as treatment planning in clinical practice. This paper proposes a modified U-Net-related architecture that integrates attention mechanisms with deep supervision for enhanced segmentation performance on the BraTS 2020 dataset. The model processes four MRI modalities and derives spatial attention to selectively highlight informative tumor regions, whereas deep supervision facilitates learning across various scales to enhance boundary accuracy as well as generalize better. Quantitative analysis shows that our model achieves a Dice coefficient of 96.20%, surpassing several emergent state-of-the-art methods. Visual inspection of predicted masks confirms accurate delineation of subregions of the tumor, such as the enhancing tumor, tumor core, as well as whole tumor, that is well aligned with expert annotations. The proposed framework achieves robust performance as well as clinical applicability. Future study entails using clinical metadata from the BraTS dataset for estimating prognosis of patients. In addition, integrating explainable AI methods (XAI) will facilitate interpreting decisions from models as well as enable transparent deployment within clinical practice settings.