CM-Oculomics fine-tuned weights
Fine-tuned DINOv2 ViT-L/14 weights (dino_deploy.pth) for predicting progression to
proliferative diabetic retinopathy during an anti-VEGF course ("anti-VEGF intolerance"
in the paper) from a single pretreatment colour fundus photograph.
This is a research prototype. It is not a medical device and must not be used to guide patient care.
Model
| Backbone | DINOv2 ViT-L/14 (Meta AI), Apache-2.0 |
| Input | one colour fundus photograph, 224 x 224 |
| Output | a model risk score in [0, 1] (not a calibrated probability: it overstates absolute risk), plus a Grad-CAM map |
| Training data | cross-validation folds 1-4 of a single-centre cohort (639 images, 57 intolerant; Longhua Hospital, Shanghai), early stopping on fold 0 (166 images); the held-out test split was never used |
| Preprocessing | identifying text burned into two cameras' images was removed outside the fundus disc before training; images are resized to 224 x 224 at inference |
| Label | progression to proliferative diabetic retinopathy after six consecutive monthly ranibizumab injections |
Reported performance
- Internal, five-fold cross-validation: patient-level AUC 0.914 (95% CI 0.870-0.951), image-level 0.952 (optimistic: early stopping used the evaluated fold)
- Held-out test split (140 images, 11 intolerant), five-seed ensemble: image-level AUC 0.930 (0.877-0.973); at the cross-validation threshold of 0.62 sensitivity was only 0.55
- Run-to-run spread over three independent training runs: image-level 0.945 +/- 0.006
- External, zero fine-tuning, DDR (proliferative vs non-proliferative proxy): AUC 0.934
Scope and limitations, please read
Performance depends strongly on the cohort. Applied without fine-tuning to seven public cohorts on three continents, external AUC ranged from 0.698 to 0.970. It varied more between cohorts than between pretraining strategies: the mean AUC of the Chinese cohorts exceeded that of the non-Chinese cohorts by 0.140, whereas the pooled difference between the two backbones was 0.054 and not consistent in direction. In a variance decomposition of the per-seed external AUCs, cohort accounted for 41%, pretraining for 25% and training seed for 35%.
We tested whether that dependence can be anticipated from unlabelled target images alone, using four published estimators of out-of-distribution performance (agreement-on-the-line in two forms, difference of confidences, average thresholded confidence) and five simpler label-free signals. None of them predicted which cohorts the model would transfer to, and the estimators were systematically optimistic: mean absolute error 0.180 to 0.244 against 0.094 for a constant equal to the mean observed accuracy (which itself needs target labels). Simply assuming no shift gave the same error as the best estimator (0.180). The estimators were applied across a change of task and class prevalence.
Validate locally before use outside the development population. About 20 labelled intolerant eyes bound balanced accuracy to within +/-0.10 and about 75-80 to within +/-0.05, whatever the site's prevalence (about 230 and 840 labelled images at 8.8%). Set the decision threshold on local data.
Further limitations: the development cohort is single-centre and retrospective, the intolerant class has 95 images, the held-out test set contains 11 proliferative events, the photographs come from three camera set-ups, and the endpoint is an outcome-anchored proxy rather than a per-injection refractoriness criterion.
Licence
Apache-2.0, consistent with the DINOv2 backbone. Free for research and commercial use with attribution.
RETFound is used in the paper only as a comparator under its CC BY-NC 4.0 licence; RETFound weights are not redistributed here.
Links
- Source code: https://github.com/23008613g/CM-Oculomics
- Archived source: https://doi.org/10.5281/zenodo.20537894
- Interactive demo: https://huggingface.co/spaces/fc28/CM-Oculomics