Research questionWhich image-text training source best transfers expert ophthalmic knowledge to vision-language models under matched data and evaluation conditions?Ophthalmic image-text training can draw on sources ranging from templated descriptions and clinical reports to broad or highly specialized literature. These sources differ in domain density and dataset composition, making it difficult to determine what drives transfer to clinical tasks. Latest papersRecent research connected to this question, newest first.Scientific Domain Knowledge Improves Vision-Language Fundus ModelsEvidence comes from identical CLIP models fine-tuned on each source, including PubMed-Ophtha: 102,023 image panels with subcaptions from 15,842 open-access PubMed Central articles. Performance is measured across 110 clinical tasks using mean linear-probing AUROC, with controls for fundus-image restriction, image-count matching, and article overlap. The evidence is limited to these models, sources, and evaluations; it suggests domain density as a likely factor but does not establish it conclusively.research paper · Sep 7, 2026