Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool, but its clinical utility is often hindered by long scan times. MRI reconstruction techniques aim to reduce scan times by enabling recovery of high quality MRI images from undersampled k-space acquisitions. This work introduces a novel diffusion model for MRI reconstruction that incorporates two key strategies: k-space spatial frequency removal and noise addition. During the forward diffusion process, k-space data points are progressively removed, simulating the undersampling process encountered in accelerated MRI scans. Simultaneously, images are corrupted via Gauss noise addition to enhance robustness against noise. This dual approach enables the diffusion model to learn the underlying image features while explicitly accounting for the acquisition process for accelerated MRI scans. Our results demonstrate that the proposed techniques offer superior image quality compared to conventional diffusion models in various undersampling rates. This work highlights that injecting prior knowledge on accelerated data acquisitions processes in MRI into diffusion models can help enhance reconstruction performance.