This work presents our solution to the BraTS Inpainting 2025 Challenge, which focuses on the reconstruction of healthy brain tissue in magnetic resonance images (MRI) where certain regions are missing or corrupted. The objective is to synthesize realistic and structurally coherent content in voided regions using only partial anatomical context. To address this, we propose a conditional denoising diffusion probabilistic model (DDPM) trained exclusively on healthy tissue. During training, we leverage all voided slices from the dataset while computing supervision only on voxels marked as healthy, ignoring regions labeled as pathological to ensure that the model learns a prior over healthy brain anatomy. For inference, we adopt a multi-stage denoising process based on DDIM (Denoising Diffusion Implicit Models), enhanced by fusion of unmasked context into each denoising stage. This strategy encourages gradual refinement and consistency with the surrounding brain structure, enabling the model to generate high-fidelity inpainting results. We evaluate our method on the challenge validation set using standard image reconstruction metrics including RMSE, PSNR, and SSIM, computed only within healthy regions. Experimental results demonstrate that our approach achieves high reconstruction accuracy and produces visually plausible results across various cases. Our findings highlight the potential of 2D conditional diffusion models, when equipped with contextual fusion, to tackle complex medical inpainting tasks in 3D volumes.