Robust simultaneous multislice MRI reconstruction using slice-wise learned generative diffusion priors.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Simultaneous multislice (SMS) imaging accelerates magnetic resonance imaging (MRI) acquisition.
• Reconstructing SMS images is challenging due to complex inter-slice signal interferences.

Purpose of the Study:

• Introduce ROGER, a robust SMS MRI reconstruction method utilizing deep generative priors.
• Enhance image quality and generalization for accelerated MRI scans.

Main Methods:

• Employ denoising diffusion probabilistic models (DDPM) for slice recovery from Gaussian noise.
• Enforce data consistency using measured k-space data within a readout concatenation framework.
• Incorporate a low-frequency enhancement (LFE) module to address autocalibration signal limitations in accelerated sequences.

Main Results:

• ROGER demonstrates superior performance compared to existing methods on both retrospective and prospective datasets.
• The method effectively enhances both anatomical and functional MRI.
• Achieved strong out-of-distribution generalization capabilities.

Conclusions:

• ROGER offers a robust and effective solution for SMS MRI reconstruction.
• The deep generative prior approach significantly improves accelerated MRI.
• The method shows promise for advancing both anatomical and functional imaging applications.