RetiGlobe

A diabetic retinopathy screening programme in Liverpool and an eye hospital in Binh Dinh photograph different populations with different cameras, so a retinal model tuned on one of them degrades on the other. The usual remedy is to fine-tune again at each site, which requires locally labelled images, a GPU and someone who can run a training loop — the exact resources a screening clinic tends not to have. RetiGlobe is built to remove that step. Its encoder stays frozen; a new site adapts the system by adding embeddings of its own reference images to the library that incoming scans are matched against, so adaptation is a database operation rather than a gradient update.

RetiGlobe is the ophthalmic foundation model developed by a consortium led by the National University of Singapore with the Joint Shantou International Eye Center of Shantou University and the Institute of High Performance Computing at A*STAR, first posted in April 2025. It is the encoder behind GlobeReady, the code-free clinician-facing platform described in the same work; the two names refer to one artifact, the model and the interface wrapped around it. Where RETFound and VisionFM are pretrained encoders intended to be fine-tuned on a downstream task, RetiGlobe is designed to be shipped fixed and queried by similarity.

#Key Features

• Training-free site adaptation: Site-specific "local feature augmentation" appends embeddings extracted from a clinic's own reference images to a global reference library, and a query inherits the label of its nearest neighbour in the combined space. Adding local features raised average diagnostic accuracy by 63.3% on fundus photographs and 17.8% on OCT across external centres.
• Two-stage pretraining across synthetic and real data: Self-supervised DINOv2 learning on 38 million synthetic ophthalmic images is followed by CLIP contrastive learning on 475,845 real image–text pairs, pairing structural detail with clinical semantics.
• Calibrated abstention: Monte Carlo dropout over 100 stochastic passes yields a confidence score per prediction, which raised accuracy on retained cases to 99.4% (fundus) and 96.2% (OCT) and supported detection of 62 out-of-distribution conditions.
• Case retrieval as explanation: Because prediction is nearest-neighbour matching, the system returns the reference images it matched, reaching Top-10 retrieval accuracy of 0.901 for fundus photographs and 0.996 for OCT.

#Technical Details

The image encoder is a ViT-L/16 backbone of roughly 307 million parameters. After DINOv2 pretraining the weights are frozen and LoRA modules of rank 8 are injected into the query and value projections, adding 0.8 million trainable parameters (0.26% of the total); image and text representations, the latter from a Bio-ClinicalBERT encoder, are aligned in a shared 512-dimensional space under a symmetric cross-entropy loss with an evidential uncertainty term. The 38 million synthetic images (25 million fundus, 13 million OCT) were produced by FundusGAN from 382,527 fundus photographs and 105,921 OCT scans drawn from 39 public datasets. Evaluation covered 488,448 images from centres in China, Singapore, Vietnam and the UK: without fine-tuning, accuracies ranged from 93.9% to 98.5% across 11 fundus diseases and 87.2% to 92.7% across 15 OCT diseases. An ablation dropping the synthetic pretraining stage cut Top-1 accuracy from 93.9% to 74.2% on fundus and from 87.2% to 56.7% on OCT.

#Applications

The intended user is an ophthalmologist or screening technician rather than an ML engineer: images are uploaded, matched and returned with a diagnosis, a confidence score and comparable reference cases. Retrieval-based inference runs on an entry- to mid-range GPU with 6–12 GB of VRAM, and because matching happens against a local reference library the system can be deployed inside a hospital network without sending patient images to an external service. In a prospective study, 31 ophthalmologists across five countries rated the platform at 83.3 on the System Usability Scale against 31.3 for Google AutoML.

#Impact

RetiGlobe reframes clinical deployment of a retinal foundation model as curating a reference library instead of retraining a network, which shifts the cost of reaching a new site from compute and ML expertise to data curation. The evidence is retrospective plus a usability questionnaire, not a prospective clinical trial, and the authors scope the model to retinal disease diagnosis: glaucoma assessment, prognostic and oculomics tasks, and biomarker quantification are outside it, and the uncertainty quantification, out-of-distribution detection and retrieval results are reported as exploratory. Under the fixed confidence threshold, 20.8% of fundus cases and 24.5% of OCT cases are flagged for clinician review. Neither code nor pretrained weights have been released: the repository carries a README stating that both will follow publication, and the GlobeReady platform is offered only under controlled non-commercial access.