Ophthalmic AI With Just Ten Examples: Comparing Multimodal Large Language Models, Vision Foundation Models, and Vision-Language Models

Ten-shot multimodal large language models (MLLMs) for retinal OCT disease classification reached a macro-F1 of 0.94, exceeding 0.59 with 10-shot vision models and matching 0.96 with full-data training.

• Why it matters: Labeled ophthalmic datasets are scarce, limiting zero-shot and few-shot AI interpretation across anterior and posterior imaging tasks and datasets.
• What they did: In a retrospective benchmark across 4 public and 3 private datasets, nine models from MLLMs, VFMs, and VLMs were tested for classification, grading, anatomical assessment, and lesion localization, using zero-shot and 10-shot settings.
• The result: Across tasks, 10-shot MLLMs also achieved QWKs of 0.69 and 0.73 for diabetic retinopathy and macular edema grading, and showed zero-shot performance on non-public clinical datasets without task-specific training.

AI summary of the abstract · may contain errors · not medical advice · read the paper.