Noise-aware adaptive diffusion sampling for accelerated knee MRI reconstruction

We present noise-aware adaptive diffusion sampling (NAD), a novel approach combining a classical noise estimation method with diffusion models for accelerated MRI reconstruction. NAD incorporates a data-consistent least-squares reconstruction as an informed starting point and uses patch-based principal component analysis to estimate the current noise level, thereby guiding adaptive sampling in the diffusion process. The method further incorporates conjugate gradient-based data consistency updates and controlled noise injection, meaning it re-injects Gaussian noise calibrated to the estimated noise levelσ^(t)and scaled byγto efficiently explore the solution space. Evaluated on the Stanford knee MRI dataset, NAD achieves higher peak signal-to-noise ratio than competing diffusion-based methods across the tested sampling budgets, with competitive or superior structural similarity index, while reducing reconstruction time in the main low-to-moderate sampling regimes. The proposed method not only advances accelerated MRI reconstruction but also provides insights into efficiently leveraging diffusion models for inverse problems in medical imaging. Code is available athere.