Understanding pre-training data effects in retinal foundation models using two large fundus cohorts - PubMed
5 hours ago
- #foundation models
- #medical AI
- #retinal imaging
- Medical foundation models pre-trained on large-scale unlabelled data show strong performance and data efficiency in clinical applications.
- The study uses two large cohorts from Moorfields Eye Hospital (UK) and the Shanghai Diabetes Prevention Program (China), each with 904,170 fundus photographs for pre-training.
- Parallel foundation models trained on individual cohorts show competitive performance on downstream tasks, even with data differing from pre-training data.
- Fairness gaps are observed over age subgroups, while sex and ethnicity show minimal impact on model performance.
- The results highlight the importance of domain-specific, fine-grained data curation for developing efficient foundation models.
- The study demonstrates the good generalisability of retinal foundation models and the varying impact of pre-training demographic attributes on fairness.