Adaptive diffusion priors for accelerated MRI reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Deep learning models for MRI reconstruction often struggle with generalization due to variations in imaging operators.
• Unconditional models offer improved reliability by decoupling image priors from specific operators, with diffusion models showing high fidelity.
• Static image priors in diffusion models can lead to suboptimal performance during inference.

Purpose of the Study:

• To develop the first adaptive diffusion prior for MRI reconstruction, named AdaDiff.
• To enhance the performance and reliability of MRI reconstruction against domain shifts.
• To improve the adaptability of diffusion models in MRI reconstruction.

Main Methods:

• AdaDiff utilizes an efficient diffusion prior trained via adversarial mapping over extended reverse diffusion steps.
• A two-phase reconstruction process is employed: an initial rapid-diffusion phase followed by an adaptation phase.
• The adaptation phase refines the reconstruction by updating the prior to minimize data-consistency loss.

Main Results:

• AdaDiff demonstrated superior performance compared to conditional and unconditional methods when subjected to domain shifts in multi-contrast brain MRI.
• The method achieved on-par or superior performance within the same domain.
• AdaDiff shows significant improvements in reliability and robustness against variable imaging operators.

Conclusions:

• AdaDiff represents a significant advancement in MRI reconstruction, offering improved adaptability and performance.
• The adaptive diffusion prior effectively addresses the limitations of static priors and domain shifts.
• This approach holds promise for more reliable and high-quality MRI acquisition across diverse clinical settings.