Artificial intelligence (AI) is transforming healthcare, particularly in diagnostics and personalized treatment. In this context, medical image segmentation is essential for delineating anatomical and pathological structures with precision. For brain tumor analysis, semantic segmentation plays a critical role in diagnosis, treatment planning, and prognosis. Although U-Net-based architectures are widely used, they face limitations in handling tumors with varying sizes, shapes, and orientations. Moreover, identifying tumor subregions increases model complexity and can lead to optimization issues, such as vanishing gradients. This study introduces UNet3D-SER, a novel 3D convolutional architecture that integrates Residual blocks (Res) and Squeeze-and-Excitation (SE) attention mechanisms to enhance feature representation and spatial context modeling. The model is evaluated on the BraTS 2020 and BraTS 2021 datasets using multimodal Magnetic Resonance Imaging (MRI) and demonstrates superior performance compared to state-of-the-art (SOTA) methods. On the BraTS 2020 validation set, our proposed architecture achieves Dice Similarity Coefficients of 0.911 for whole tumor (WT), 0.866 for tumor core (TC), and 0.817 for enhancing tumor (ET). On BraTS 2021, performance improves further with Dice scores of 0.94 (WT), 0.94 (TC), and 0.89 (ET). These results confirm the model’s robustness in capturing tumor heterogeneity and its strong potential for clinical deployment in brain tumor segmentation tasks.