RFM: A Lightweight Retinal Foundation Model for Generalised Oculomics
- 1School of Computer Science and Mathematics, Kingston University, London, UK
Abstract
Foundation models for colour fundus photography (CFP) are candidate visual encoders for multimodal medical AI, but existing models are large and are assessed under fine-tuning rather than the frozen-encoder conditions such systems impose. We present RFM (Retinal Foundation Model), a 64-million-parameter Vision Transformer initialised from DINOv3-Base and domain-adapted with the LeJEPA joint-embedding predictive self-supervised objective with domain-specific pretext tasks, on 4.3 million CFPs from the North East London Diabetic Eye Screening Programme. A single linear head per task on one frozen backbone measures representation generalisability, not fine-tuning capacity. Across six external datasets spanning diabetic retinopathy (DR) grading, glaucoma detection, ten UK Biobank biomarker phenotypes and 3-year incidence of myocardial infarction, ischaemic stroke and Parkinsons disease, RFM attains the lowest mean absolute error of any frozen model on all seven regression phenotypes and a higher area under the receiver operating characteristic curve (AUROC) than any RETFound variant on all eight matched classification benchmarks. It also attains the highest AUROC of any model evaluated on four of the five ophthalmic benchmarks, with 4.8 x fewer parameters than RETFound-MAE and RETFound-DINOv2. Applied to UK Biobank without retraining, RFM exceeds the best published retinal foundation-model AUROC for 3-year myocardial infarction and ischaemic stroke; the Parkinsons endpoint (11 cases) is inconclusive. DINOv3-Base without retinal pre-training, read the same way, stays competitive on DR grading but falls clearly behind on glaucoma, and the feature-extraction convention alone is worth up to 0.154 AUROC. One lightweight backbone yields generalised representations serving ophthalmic grading, biomarker regression and incidence prediction.
Related
- Integrated oculomics and lipidomics characterise cross-sectional retinal–lipidomic associations in a generally healthy biobank population
- Atrial fibrillation and retinal imaging-based oculomics: A cross-sectional and longitudinal analysis of two large cohorts
- Explainable foundation model for dementia screening and risk stratification using retinal fundus images
- Representation Transfer of Foundation Models for Ultra-Widefield Retinal Imaging.
- Prediction of incident atrial fibrillation from retinal fundus images using a multimodal foundation model
BibTeX
@misc{fajtl2026lightweight,
title = {{RFM: A Lightweight Retinal Foundation Model for Generalised Oculomics}},
author = {J. Fajtl and R. A. Welikala and R. Shakespeare and R. Chambers and L. Bolter and J. Anderson and P. Remagnino and A. P. Menon and G. Johnson and F. Rahman and A. Olvera-Barrios and A. Warwick and C. Egan and A. Tufail and C. G. Owen and A. R. Rudnicka and S. Barman},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.02.26362554},
url = {https://www.medrxiv.org/content/10.64898/2026.10.02.26362554v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - RFM: A Lightweight Retinal Foundation Model for Generalised Oculomics
AU - J. Fajtl
AU - R. A. Welikala
AU - R. Shakespeare
AU - R. Chambers
AU - L. Bolter
AU - J. Anderson
AU - P. Remagnino
AU - A. P. Menon
AU - G. Johnson
AU - F. Rahman
AU - A. Olvera-Barrios
AU - A. Warwick
AU - C. Egan
AU - A. Tufail
AU - C. G. Owen
AU - A. R. Rudnicka
AU - S. Barman
PY - 2026
DA - 2026/10/02
PB - medRxiv
DO - 10.64898/2026.10.02.26362554
UR - https://www.medrxiv.org/content/10.64898/2026.10.02.26362554v1
AB - Foundation models for colour fundus photography (CFP) are candidate visual encoders for multimodal medical AI, but existing models are large and are assessed under fine-tuning rather than the frozen-encoder conditions such systems impose. We present RFM (Retinal Foundation Model), a 64-million-parameter Vision Transformer initialised from DINOv3-Base and domain-adapted with the LeJEPA joint-embedding predictive self-supervised objective with domain-specific pretext tasks, on 4.3 million CFPs from the North East London Diabetic Eye Screening Programme. A single linear head per task on one frozen backbone measures representation generalisability, not fine-tuning capacity. Across six external datasets spanning diabetic retinopathy (DR) grading, glaucoma detection, ten UK Biobank biomarker phenotypes and 3-year incidence of myocardial infarction, ischaemic stroke and Parkinsons disease, RFM attains the lowest mean absolute error of any frozen model on all seven regression phenotypes and a higher area under the receiver operating characteristic curve (AUROC) than any RETFound variant on all eight matched classification benchmarks. It also attains the highest AUROC of any model evaluated on four of the five ophthalmic benchmarks, with 4.8 x fewer parameters than RETFound-MAE and RETFound-DINOv2. Applied to UK Biobank without retraining, RFM exceeds the best published retinal foundation-model AUROC for 3-year myocardial infarction and ischaemic stroke; the Parkinsons endpoint (11 cases) is inconclusive. DINOv3-Base without retinal pre-training, read the same way, stays competitive on DR grading but falls clearly behind on glaucoma, and the feature-extraction convention alone is worth up to 0.154 AUROC. One lightweight backbone yields generalised representations serving ophthalmic grading, biomarker regression and incidence prediction.
ER -