Recurrent retinal detachment demo — image extractor
One file: extractor.enc, 1157 MB. It holds the image extractor for BartvanderWoude/recurrent-RD-demo in
the packaged form that demo reads, and it is not a model file you can load on
your own - from_pretrained and onnxruntime will both refuse it. It is of no
use outside the demo.
Image extractor
Image features come from RETFound, by Yukun Zhou and colleagues (https://github.com/rmaphoh/RETFound_MAE), described in A foundation model for generalizable disease detection from retinal images, Nature 622, 156-163 (2023).
RETFound is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0): https://creativecommons.org/licenses/by-nc/4.0/. This demo is a non-commercial research demonstration.
Modified: the published weights were exported to ONNX for the browser at their original fp32 precision. They were not retrained or fine-tuned; the run this demo comes from used the released weights frozen, as a fixed feature extractor.
Hosting requirements
This repository is not a mirror; it is what the demo actually fetches, so three things about it are load-bearing.
- It has to stay public. The page fetches it cross-origin with no credentials and no token - a static page has nowhere to keep one. A private or gated repo cannot be made to work.
- The file has to stay at the root under this exact name. The page appends
extractor.encto the base URL it was built with. - Replacing it invalidates the demo, which is built against these exact bytes. Publish a new file and a rebuilt demo together, or neither.