CLEAR-EC Phase II: tier p3b checkpoints
Weights of the tier p3b model for the CLEAR-EC Grand Challenge, Phase II. Given one corneal endothelium specular microscopy image (972 x 1296, 8-bit), the model predicts endothelial cell density (CD), the coefficient of variation of cell area (CV) and hexagonality (HEX).
Code, container definition and run instructions: the GitHub repository leebyounghan/clear-ec-research (access on request).
Contents
443 files, 16.25 GB, at the relative paths the container expects. checkpoint_sha256.txt lists sha256, size and path of every file.
The container mounts this folder read-only at /opt/ml/model. The --include patterns keep this card and the hash list out of
checkpoints/, so the code repository's own checkpoints/README.md is not overwritten.
hf download Byounghan/clear-ec-p3b --local-dir checkpoints/ --include "cpsam_models/*" --include "experiments/*"
python3 scripts/verify_checkpoints.py
Results (mean absolute percentage error)
| Mean | CD | CV | HEX | |
|---|---|---|---|---|
| Out-of-fold, five donor-grouped folds | 8.660 | 6.148 | 10.112 | 9.719 |
| Phase II hidden test set (1,000 images) | 8.935 | 6.162 | 10.290 | 10.353 |
Notes
- Some regression checkpoints are stored in fp16 and are cast back to fp32 at load time (list in the code repository).
- The five files under
experiments/lane_cd_20261108/arms/dinov3_b16_f99_s47are not read by p3b; they are kept so the tree matches the tested archive. - Trained on the CLEAR-EC challenge training data and the public Topcon specular set. Research use only; not for clinical use.
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support