Dual-Domain Multipath Self-Supervised Diffusion Model for Accelerated MRI Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Magnetic Resonance Imaging (MRI) acquisition is time-consuming, limiting clinical efficiency and patient comfort.
• Deep learning, especially diffusion models, shows promise for accelerating MRI reconstruction.
• Current diffusion models face challenges with training data requirements, computational cost, and lack of uncertainty estimation.

Purpose of the Study:

• To develop a novel framework, DMSM, for enhanced accelerated MRI reconstruction.
• To address limitations of existing diffusion models, including data dependency, computational burden, and absence of uncertainty quantification.
• To improve reconstruction accuracy, efficiency, and clinical explainability in accelerated MRI.

Main Methods:

• Proposed a dual-domain multipath self-supervised diffusion model (DMSM).
• Integrated a self-supervised dual-domain diffusion model training scheme.
• Employed a lightweight hybrid attention network (LHAN) and a multipath inference strategy.

Main Results:

• DMSM demonstrated superior performance compared to supervised and self-supervised baselines on human MRI datasets.
• The model effectively preserved fine anatomical structures and reduced artifacts at high acceleration factors.
• Generated uncertainty maps correlated with reconstruction errors, offering clinical interpretability.

Conclusions:

• DMSM offers a practical and effective solution for accelerated MRI reconstruction, overcoming key limitations of prior diffusion models.
• The framework enhances diagnostic confidence through uncertainty estimation.
• DMSM is a significant advancement for efficient and reliable clinical MRI.