Our study tackles the complex task of segmenting multiple classes of brain tumors using contrast-enhanced MRI scans. A dataset comprising 3,064 T1-weighted images from 233 patients was employed to evaluate the efficacy of five deep learning architectures: U-Net, SegNet, V-Net, U-Net augmented by ResNet, and LadderNet. These models were further optimized using Attention Gates to concentrate on tumor regions, Swin Transformers to extract hierarchical features, and Denoising Generative Adversarial Networks (DGANs) to mitigate noise in the MRI scans. Preprocessing steps—such as normalization, contrast enhancement, skull stripping, and data augmentation—were implemented to address challenges posed by variations in tumor size and intensity. The performance of each model was assessed using metrics including the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), allowing for a clear comparison. The findings highlight the promise of deep learning methods for automated brain tumor detection and suggest a valuable framework for early diagnosis, ultimately advancing patient outcomes.