In medical imaging, precise identification of brain tumors in MRI is crucial; however, it remains a challenge. While manual segmentation suffers from time constraints and observer variability, automated approaches using deep learning have emerged as promising alternatives. This research evaluates four contemporary deep learning architectures namely U-Net, Attention U-Net, YOLOv8, and SAM (Segment Anything Model) to assess their performance. It utilizes a comprehensive dataset of 3,064 MRI scans with corresponding expert-annotated tumor masks. Performance assessment relies on two key metrics: the Dice Similarity Coefficient (DSC) for segmentation accuracy and mean Average Precision (mAP) for detection reliability. Initial results demonstrate that Attention U-Net exhibits superior performance with a mean DSC of 0.837, outperforming other models across multiple evaluation criteria. This comparative analysis not only emphsise the relative strengths and limitations of each architecture but also offers meaningful perspectives for their practical implementation in clinical settings. The findings aim to guide both research directions and clinical applications for identifying and analysing brain tumors.