In Part 2 of this year’s round-up of landmark literature, The Ophthalmologist Power Lister Andrzej Grzybowski reviews the key artificial intelligence (AI) research papers from 2024-25.
AI: Foundation models in ophthalmology
A foundation model is a large, general-purpose AI model that is trained on a vast and diverse dataset using self-supervised or unsupervised learning. Once pretrained, it can be adapted (or "fine-tuned") for a wide range of downstream tasks – often with minimal additional labeled data. Foundation models serve as universal starting points for building specialized AI applications. In medical fields like ophthalmology, foundation models:
Reduce the need for large, labeled clinical datasets.
Improve generalization across patient populations, devices, and imaging settings.
Enable multitask and multimodal workflows (e.g., combining retinal images with clinical notes or demographics).
Provide building blocks for safer, more scalable AI systems (e.g., RETFound for retinal disease).
Y Zhou et al., “A foundation model for generalizable disease detection from retinal images,” Nature, 622, 156 (2023). PMID: 37704728.
This seminal study, published in 2023, introduced the first foundation model to ophthalmology, the RETFound model, based on self-supervised masked autoencoders and pretrained on 1.6 million unlabeled retinal images. It outperformed traditional ImageNet models in diagnosis, prognosis, and systemic disease prediction, requiring fewer labeled samples. RETFound marked the shift to foundation models for retinal disease detection.
In 2024, several important studies showed new applications and performance of foundation models in ophthalmology (see Table 1).
J Zhang et al., “RETFound-enhanced community-based fundus disease screening: real-world evidence and decision curve analysis,” NPJ Digit Med., 7, 108 (2024). PMID: 38693205.
RETFound was adapted to real-world community screening in China, showing >15 percent better sensitivity and specificity compared to commercial AI tools. The model demonstrated high resilience to poor imaging conditions. Decision curve analysis confirmed a significant net benefit for both rural and urban screening.
C Nielsen et al., “Foundation model-driven distributed learning for enhanced retinal age prediction,” J Am Med Inform Assoc., 31, 2550 (2024). PMID: 39225790.
Using a compressed RETFound version, a distributed learning system predicted retinal age gaps. Federated and traveling model learning achieved equivalent accuracy to centralized training but were more computationally efficient.
K Du et al., “Detection of Disease Features on Retinal OCT Scans Using RETFound,” Bioengineering (Basel), 11, 1186 (2024). PMID: 39768004.
This study evaluated RETFound, a foundation model pretrained on OCT images, for automated classification of retinal disease features, comparing its performance to ResNet-50 using a labeled dataset. RETFound achieved comparable accuracy and AUC-ROC values, demonstrating its potential to support more efficient and consistent OCT image interpretation in clinical settings.
MS Chen et al., “Independent Evaluation of RETFound Foundation Model's Performance on Optic Nerve Analysis Using Fundus Photography,” Ophthalmol Sci., 28, 100720 (2025). PMID: 40161459.
This study assessed RETFound for predicting optic nerve metrics like cup-to-disc ratio (CDR) and RNFL thickness from fundus images. RETFound outperformed VGG16 and ViT feature extractors, achieving high R² values (up to 0.961) in single-output tasks, though performance was lower in multioutput predictions. These results highlight RETFound's potential to enable accurate optic nerve evaluation from fundus photos, even without task-specific training.
D Kuo et al., “How Foundational Is the Retina Foundation Model? Estimating RETFound's Label Efficiency on Binary Classification of Normal versus Abnormal OCT Images,” Ophthalmol Sci., 5, 100707 (2025). PMID: 40161460.
This study evaluated the label efficiency of the RETFound model for classifying normal vs. abnormal OCT B-scans in diabetic retinopathy screening. RETFound consistently outperformed ResNet-50 and standard ViT models across all dataset sizes, particularly excelling with limited training data. These results highlight the value of retina-specific pretraining and suggest RETFound's strong potential for scalable, label-efficient ophthalmic diagnostics.
B Chuter et al., “Evaluating a Foundation Artificial Intelligence Model for Glaucoma Detection Using Color Fundus Photographs,” Ophthalmol Sci., 5, 100623 (2024). PMID: 39650567.
This study evaluated RETFound’s performance in detecting glaucoma from optic disc photographs across varying training sizes, epochs, and patient subgroups. RETFound achieved strong accuracy (AUC up to 0.86), with performance improving with more data and training, and remaining consistent across age and race. Its adaptability and effectiveness, even with limited data, highlight RETFound’s potential for scalable, automated glaucoma detection in diverse clinical settings.
T Lin et al., “Efficiency and safety of automated label cleaning on multimodal retinal images,” NPJ Digit Med.. 5, 10 (2025). PMID: 39757295.
This study evaluated automated label cleaning using Cleanlab in noisy fundus and OCT datasets, demonstrating substantial improvements in label accuracy (up to 62.9 percent) and dataset quality. RETFound’s classification accuracy improved significantly (up to 52.9 percent) when trained on cleaned data, with most label errors accurately corrected and minimal over-cleaning using a DQS-guided strategy. These findings highlight the effectiveness and safety of automated label correction in improving model training and dataset reliability.
IN 2024 AND 2025 NEW FOUNDATION MODELS WERE DEVELOPED (SEE TABLE 2).
J Qui et al., “Development and validation of a multimodal multitask vision foundation model for generalist ophthalmic artificial intelligence,” NEJM AI, 1 (2024). DOI: 10.1056/ AIoa2300221.
VisionFM is a multipurpose ophthalmic foundation model pretrained on 3.4 million images spanning eight imaging modalities and diverse diseases, designed to generalize across clinical tasks. It outperformed baseline models in internal and external validations, achieving AUROCs up to 0.974 and showing diagnostic accuracy comparable to intermediate-level ophthalmologists. VisionFM demonstrated strong generalizability to unseen modalities and tasks, supporting its potential as a scalable, open-source platform for broad ophthalmic AI applications.
M Baharoon et al., “HyMNet: A multimodal deep learning system for hypertension prediction using fundus images and cardiometabolic risk factors,” Bioengineering, 11, 1080 (2024). PMID: 39593740.
HyMNet combined RETFound-based imaging features and demographic factors for hypertension detection. It achieved an F1 score of 0.771, especially improving prediction among diabetic patients.
Z Zhang et al., “Effective automatic classification methods via deep learning for myopic maculopathy,” Front Med (Lausanne). 11:1492808 (2024). PMID: 39606624.
An ensemble deep learning system using RETFound, ViT, and ResNet achieved 95.4 percent accuracy in classifying myopic maculopathy. The system demonstrated excellent performance in complex retinal disease detection tasks.
Y Peng et al., “Enhancing AI reliability: A foundation model with uncertainty estimation for optical coherence tomography-based retinal disease diagnosis,” Cell Rep Med., 6, 101876 (2025). PMID: 39706192.
A foundation model with uncertainty estimation (FMUE) to detect 16 retinal conditions on optical coherence tomography (OCT) was developed. In the internal test set, FMUE achieved a higher F1 score of 95.74 percent than other state-of-the-art algorithms (92.03 percent-93.66 percent) and improved to 97.44 percent with threshold strategy. In human-model comparison, FMUE achieved a higher F1 score of 96.30 percent than retinal experts (86.95 percent, p = 0.004), senior doctors (82.71 percent, p < 0.001), junior doctors (66.55 percent, p < 0.001), and generative pretrained transformer 4 with vision (GPT-4V) (32.39 percent, p < 0.001).
