In a retrospective comparative study, RETFound—a retina-specific self-supervised foundation model—matched traditional deep-learning models on full datasets but improved low-label glaucoma detection, reaching an AUC of 0.908 with 400 training images and outperforming ResNet50 on internal and external test sets. The study also found advantages for RETFound in diabetic retinopathy and systemic disease detection when fine-tuned on smaller datasets, based on testing across SEED, APTOS 2019, and multiple external cohorts.

Why It Matters To Your Practice

• AI performance may depend less on model type when large labeled ophthalmic datasets are available, but foundation models can help when labels are scarce.

• This is especially relevant for glaucoma screening workflows, where assembling high-quality labeled images is often difficult and expensive.

• For clinicians evaluating vendor tools, the study suggests that pretraining strategy matters most in low-data settings rather than fully resourced ones.

Clinical Implications

• For glaucoma, RETFound achieved an internal-test AUC of 0.908 after training on 400 images, with external AUCs of 0.835 on ODIR-5K, 0.779 on PAPILA, and 0.990 on GAMMA.

• When full datasets were used, RETFound and conventional models performed similarly for ocular disease detection overall.

• ResNet50 was inferior to RETFound for diabetic retinopathy at very small sample sizes and for glaucoma at 400 images or fewer, while SwinV2 remained broadly comparable to RETFound.

Insights

• RETFound is a retina-specific foundation model trained with self-supervision, which may explain its relative strength when labeled examples are limited.

• The study compared RETFound with ResNet50, ViT-Base, and SwinV2 across ocular and systemic diseases, using both internal and external validation datasets.

• Generalizability looked encouraging but not uniform across external glaucoma datasets, underscoring the need for local validation before deployment.

The Bottom Line

• If your practice or health system has limited labeled retinal data, RETFound-like foundation models may offer a practical edge for glaucoma and some other detection tasks.

• If you already have large, well-annotated datasets, standard deep-learning models may perform just as well for many ocular applications.

• Clinicians should view this as a signal to ask AI vendors not just how accurate a model is, but how much labeled data it needed to get there.