Magnetic Resonance Imaging (MRI) scans are limited by long acquisition times required for Nyquist-rate sampling, leading to motion artefacts and high costs. This work develops an end-to-end diffusion-based MRI denoising framework to enhance image quality under undersampled conditions. A modified 2D U-Net diffusion model was adapted for complex-valued MRI data, trained using a custom multi-metric loss balancing perceptual and quantitative fidelity. The system efficiently processes large slices via patch-based denoising and achieves stable convergence with improved SSIM and PSNR. The trained denoiser is validated as a standalone model and will next be integrated as a plug-and-play proximal operator in MRI reconstruction.