Mutli-modal straight flow matching for accelerated MR imaging.

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Science

Background:

• Diffusion models are increasingly used for Magnetic Resonance (MR) image reconstruction.
• Iterative denoising in diffusion models requires thousands of steps, leading to high computational complexity and limited applications.
• Existing methods struggle with balancing image quality and computational cost in MR reconstruction.

Purpose of the Study:

• To develop a novel generative model for efficient and high-quality MR image reconstruction.
• To reduce the number of inference steps and computational complexity in MR image reconstruction.
• To leverage multi-modal data for enhanced reconstruction of target MR modalities.

Main Methods:

• Introduced a novel straight flow matching model based on a neural ordinary differential equation (ODE) generative model.
• Developed a multi-modal straight flow matching (MMSflow) model incorporating low and high-frequency fusion layers.
• Utilized easily available MR modalities to guide the reconstruction of target modalities.

Main Results:

• The proposed model creates a linear path between undersampled and reconstructed images, simulated with few Euler steps.
• MMSflow achieved state-of-the-art performance in reconstruction tasks on fastMRI and Brats-2020 datasets.
• The method improved the sampling rate by an order of magnitude compared to stochastic differential equation (SDE) based methods.

Conclusions:

• Straight flow matching offers a computationally efficient alternative for MR image reconstruction.
• Multi-modal data fusion with specialized layers enhances reconstruction quality and efficiency.
• The MMSflow model represents a significant advancement in accelerating MR image acquisition and reconstruction.