hwang714/retivasc-tpot

RetiVASC · TPOT enhancement (ONNX)

The fundus photograph enhancement that the RetiVASC pipeline of RetiVASC runs in the web browser before it measures the vessels. Both files come from the lab's TPOT-SR repository (commit 0620afa), exported to ONNX by RetiVASC's tools/export_tpot.py. The ONNX files match PyTorch to within 4e-6.

tpot.onnx      TPOT enhancement, one pass at 256 x 256 (weights/best_SSIM.pth)            19 MB, float32
fusion.onnx    TPOT-SR fusion head, restores the detail at 2048 x 2048, run in 512 tiles  0.5 MB, float32
manifest.json  both files with their size and SHA-256

Each release has a tag (v0.1.0 now), and RetiVASC loads a tag, never main. To use these in your own build of RetiVASC, set VITE_TPOT_URL=https://huggingface.co/hwang714/retivasc-tpot/resolve/v0.1.0/.

Licence and attribution

TPOT and TPOT-SR are models of the Geometry Systems Lab, Arizona State University (Retinal-Research). Their source repository does not state a licence yet. Until it does, these weights are shared for research use with RetiVASC; contact the authors for any other use. Paper: arXiv:2411.01403.

These models are for research use. They are not a medical device and must not be used for diagnosis.