Health system learning enables generalist neuroimaging models

Authors/institution: A. Kondepudi and collaborators across health-system and academic institutions. Venue: Nature Medicine, peer reviewed. Problem: medical models trained on curated or internet data may not capture routine clinical complexity. Contribution: NeuroVFM, trained on 5.24 million CT/MRI series, with diagnosis, triage and report-generation evaluations. Practical significance: domain foundation models could become auditable tools used by broader clinical agents. Limitations: prospective workflow impact, fairness, uncertainty and multi-system generalisation remain open. Paper.