A new AI architecture built for medical imaging can spot rare diseases that other systems miss entirely.
The approach, called Generalist-Specialist Mixture-of-Experts (GS-MoE), is described in an arXiv preprint (arXiv:2609.18688) posted September 17, 2026, and has not yet been peer-reviewed. It splits the model into two branches: a "generalist" that learns shared patterns across imaging modalities like X-ray, CT, and ultrasound, and modality-specific "specialist" experts, fused together with constraints designed to keep cross-modal signal from getting lost. Tested on RadImageNet, a benchmark of 1.35 million images covering 165 pathologies across three modalities, GS-MoE detected six low-prevalence conditions that every baseline model scored a flat zero on, with per-class F1 gains as high as 0.60. It also matched or slightly beat dense and specialist-only baselines overall (MCC 0.770) while using about 53% fewer active parameters at inference.
Most medical-imaging AI is trained and scored on common conditions, because that's where the data is. Rare pathologies get buried, and a model that scores well on average can still be clinically useless for the cases doctors most need help catching. If the zero-to-nonzero jump on six rare classes holds up, that's a bigger deal than the aggregate score, and doing it with fewer parameters suggests the gain isn't just "throw more compute at it."
Still, this is one un-peer-reviewed preprint on one benchmark. RadImageNet performance is not the same as clinical performance, and "recovers detection" from F1=0 doesn't tell us how reliable those rare-case predictions actually are.