Dynamic cardiac MRI requires rapid image acquisition to freeze heart motion. Accelerated scanning undersamples the data, creating aliasing artifacts that must be resolved computationally. Standard methods — compressed sensing, neural networks — trade reconstruction time for image quality.
The authors (arXiv:2603.21980) repurpose Gaussian splatting, originally developed for 3D scene rendering, as a cardiac MRI reconstruction method. Multiple Gaussian primitives are trained to represent tissue properties, with the splatting framework extended to model anatomical motion. The result: high-quality cardiac images from 16× undersampled data, with training completing in 60 seconds and inference in 0.5 milliseconds.
The unexpected finding: the properties of the learned Gaussians — their positions, widths, and intensities — directly encode physiological properties of the heart. The representation isn't just a compression trick; it maps to anatomy.
The through-claim: explicit representations outperform implicit ones when the underlying structure has natural decomposition into discrete objects. The heart is made of distinct tissue regions, and Gaussian splatting's representation — a set of localized, parameterized components — matches that structure. The model succeeds not by brute-force learning but by structural correspondence between the representation and the anatomy.