AtlasGS uses Gaussian splatting to upscale brain MRIs 7x across modalities

A new paper from researchers led by Yifan Gao introduces AtlasGS, a method that uses Gaussian splatting—a 3D rendering technique—to dramatically improve the spatial resolution of brain MRIs across multiple imaging modalities. Clinical MRI often suffers from inconsistent resolutions: high-quality structural scans (e.g., T1-weighted) are acquired with thin slices, but functional or pathological sequences like FLAIR, DWI, and ASL are typically collected with thicker, sparser slices. AtlasGS solves this by first learning a shared Gaussian geometry scaffold from a high-resolution isotropic scan, capturing the subject’s precise brain anatomy. In a second stage, this scaffold is reused to guide the reconstruction of any target modality from its low-resolution sparse slices, effectively performing through-plane super-resolution at factors of 3x, 5x, and even 7x.

The method was validated on three large, diverse datasets: UK Biobank (healthy subjects), GBM (glioblastoma patients with tumors), and ABCD (pediatric population). AtlasGS consistently outperformed existing super-resolution techniques across all modalities, including T2-weighted, FLAIR, DWI, and ASL perfusion images. Beyond simple upscaling, the shared geometry enables arbitrary-view synthesis—radiologists can reconstruct sagittal, coronal, or oblique slices from a single low-resolution acquisition—without loss of structural consistency. The framework also shows promise for self-supervised in-plane super-resolution, which could further reduce scan times. By bridging resolution gaps between imaging protocols and offering interpretable geometric priors, AtlasGS establishes a new paradigm for retrospective MRI harmonization, potentially allowing multi-center studies and clinical tools to work reliably across different scanners and sequences.