Gaussian primitives beat implicit neural networks in medical imaging

A new arXiv paper by Stolt-Ansó, Dannecker, Huang, Jakab, and Rueckert argues that explicit Gaussian primitives should replace implicit neural representations (INRs) for medical image reconstruction. INRs rely on global multi-layer perceptrons (MLPs) to encode continuous, resolution-independent images, but they require heavy compute, large memory, and long optimization times. As medical imaging moves toward ever-higher-resolution volumes, these bottlenecks become critical. The authors propose switching to localized, rasterization-friendly Gaussian primitives—the same idea behind 3D Gaussian splatting—which avoid deep network evaluations by storing explicit, position-aware primitives.

The paper backs this claim with a cross-dimensional benchmark: 2D microscopy histology and 3D lung CT scans. Across all compression factors tested, Gaussian representations matched or exceeded the reconstruction fidelity (PSNR/SSIM) of implicit methods, while showing significantly lower optimization times and memory footprints. The authors also outline mathematical advantages of explicit primitives, including better locality and interpolation properties, which enable faster training and more efficient encoding of sharp boundaries typical in medical structures. The results strongly suggest that Gaussian primitives are a practical, scalable alternative to INRs for high-resolution medical imaging, opening the door to real-time processing of large volumes on standard hardware.