Deep learning plays an important role in improving brain tumor segregation for accurate diagnosis and treatment planning, which helps medical image analysis. Brain tumor segmentation from MRI scans remains challenging due to the need for precise delineation between tumor boundaries and healthy tissue, which is essential for effective patient care. This research specifically examines the efficacy of convolutional neural network (CNN) architectures, with a focus on U-Net and three-dimensional CNNs, which are known for their success in segmenting complex structures in MRI images. We address the central research question: How can the integration of Feature Pyramid Networks (FPN) with U-Net improve segmentation performance in identifying regions of brain tumors? This work explores both the capabilities and limitations of existing CNN-based models, such as the need for large annotated datasets, varying imaging quality, and high computational demands. Additionally, it addresses challenges related to model interpretability, overfitting, and ethical considerations in deploying AI in healthcare.