Noise-aware adaptive diffusion sampling for accelerated knee MRI reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Accelerated MRI reconstruction is crucial for reducing scan times and improving patient comfort.
• Diffusion models show promise for medical image reconstruction but often require significant computational resources.
• Existing methods lack efficient strategies for noise handling and adaptive sampling during reconstruction.

Purpose of the Study:

• To introduce Noise-Aware adaptive Diffusion sampling (NAD), a novel method for accelerated MRI reconstruction.
• To enhance the efficiency and performance of diffusion models in medical imaging inverse problems.
• To reduce computational time while improving image quality in accelerated MRI.

Main Methods:

• NAD combines classical noise estimation (patch-based PCA) with diffusion models for informed initialization.
• It employs conjugate gradient-based data consistency updates and controlled Gaussian noise injection.
• Adaptive sampling is guided by estimated noise levels ($\hat{\sigma}(t)$) and a scaling factor ($\gamma$).

Main Results:

• NAD consistently outperforms state-of-the-art diffusion-based methods on the Stanford Knee MRI dataset.
• Significant reductions in computational time were achieved compared to existing techniques.
• Improvements in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) were observed.

Conclusions:

• NAD offers a significant advancement in accelerated MRI reconstruction.
• The method demonstrates the potential of diffusion models for efficient inverse problems in medical imaging.
• NAD provides a framework for leveraging noise estimation to guide adaptive sampling in diffusion-based reconstruction.