Clinical brain MRI scans for Alzheimer's disease diagnosis often suffer from motion artifacts, signal dropout, and incomplete acquisitions. While deep learning methods like LaMa produce visually coherent inpainting, they may alter critical anatomical structures, and traditional methods such as OpenCV Telea maintain local continuity but introduce excessive smoothing. This study presents a dual-stage generative adversarial network with attention gates for medical image restoration, uniquely complemented by a novel gradient-weighted hybrid blending strategy that adaptively combines GAN outputs with classical inpainting based on local image gradients. The architecture employs an attention-enhanced U-Net generator and refinement network, supervised by patch and global discriminators through combined adversarial, reconstruction, perceptual, structural similarity, and gradient losses. Evaluation on 12,491 balanced Alzheimer's MRI scans across four severity stages with synthetic 10–30% occlusions shows that the proposed GAN reduces mean squared error by 48.7% versus LaMa and 41.3% versus OpenCV, achieving 17.44 dB PSNR and 0.9796 SSIM (computed over the brain region). The gradient-guided Hybrid-GAN maintains reconstruction quality with a total inference time of approximately 8.97 ms per image. Downstream VGG16 classification reveals Hybrid-GAN preserves diagnostic information most effectively, attaining 96.05% accuracy, only 0.96 points below the 97.01% baseline and surpassing all alternative methods. These results demonstrate that attention-driven generative models combined with structure-aware blending provide a practical and novel solution for artifact mitigation in neuroimaging diagnostics.