Fast undersampled dynamic MRI reconstruction using explicit representation learning with Gaussian splatting

TL;DR

This paper introduces a fast, explicit representation learning method using Gaussian splatting for high-quality dynamic MRI reconstruction, enabling rapid training and inference with physiological encoding.

Contribution

It presents a novel Gaussian splatting framework for dynamic MRI that models tissue and motion explicitly, achieving rapid training and high-quality results.

Findings

Gaussian splats trained in 60 seconds

Achieves high-quality cardiac MRI at R=16

Gaussian properties encode physiological information

Abstract

Motivation: Quickly obtaining high-quality MRI from accelerated acquisitions is important to mitigate motion artifacts, maintain patient comfort, and improve clinical efficiency. Goals: To obtain high-quality dynamic MRI using efficient, personalized models. Approach: We propose a novel explicit representation learning approach using Gaussian splatting. Multiple Gaussian primitives are trained to represent the underlying tissue. We extend the Gaussian splatting framework to model anatomical motion, enabling learning an efficient, explicit representation of dynamic MRI. Results: Gaussian splats can be trained in 60s with 0.5ms/dynamic inference time. High-quality cardiac MRI is obtained at R=16. We show that the properties of the Gaussians directly encode physiological properties.