Magnetic resonance imaging (MRI) scans are slow. AI can speed them up by reconstructing images from less data, but these models fail on new scanners, anatomies, or hospitals because training on enough varied real medical scans is expensive and blocked by privacy rules.
The authors built SIM (Scalable Inverse Model), which generates its own training data by simulating the physics of MRI acquisition on synthetic objects. They paired this with a high-capacity network designed to exploit massive compute, replacing the need for limited real data.
SIM generalized to real scans it had never seen, outperforming standard AI methods across 14 diverse datasets. Crucially, its performance scales with compute, mirroring breakthroughs in language models. This establishes a scalable, privacy-preserving path to universal AI for medical imaging.