RetiVASC · AutoMorph web models (ONNX)
These are the AutoMorph models that RetiVASC runs in the web browser (WebGPU, or WebAssembly when there is no WebGPU). They are the ensembles of AutoMorph, exported to ONNX from the AutoMorph v2 runtime checkpoints. RetiVASC downloads them once, checks each file against the SHA-256 in its manifest.json, and caches them in the browser.
quality/ quality_0.onnx … quality_7.onnx M1 image quality (EfficientNet-B4), 8 members
vessel/ vessel_0.onnx … vessel_9.onnx M2 binary vessel segmentation (U-Net), 10 members
artery_vein/ artery_vein_0.onnx … artery_vein_7.onnx M2 artery/vein (BF-Net), 8 members
disc_cup/ disc_cup.onnx M2 optic disc and cup (W-Net, 8 members in one graph)
manifest.json every file with its size and SHA-256; the ensembles are averaged in AutoMorph v2's order
Each release of the models has a tag (v0.1.0 now), and RetiVASC loads a tag, never main. A deployed site therefore keeps its models when new ones are added here. To load them from your own build of RetiVASC, set VITE_MODELS_URL=https://huggingface.co/hwang714/retivasc-automorph/resolve/v0.1.0/.
Precision: the weights are stored as float16 and the arithmetic runs in float32. The disc/cup model is float32 throughout, because it sets the M3 measurement zones. The 72 vascular features computed from these models in the browser agree with the original Python AutoMorph to within 0.4%. See the RetiVASC README for the full comparison.
Licence and attribution
The weights are AutoMorph's, redistributed under its Apache License 2.0. They are not new models and were not retrained. If you use them, please cite AutoMorph:
Zhou Y, Wagner SK, Chia MA, et al. AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline. Translational Vision Science & Technology 11(7):12, 2022. https://doi.org/10.1167/tvst.11.7.12
These models are for research use. They are not a medical device and must not be used for diagnosis.