clear-ec-phase2 / README.md
RINTHIRA's picture
Model card: team name PRL CorneaForge
a01ab25 verified
|
Raw History Blame Contribute Delete
2.49 kB
metadata
tags:
  - medical-imaging
  - regression
  - corneal-endothelium
  - specular-microscopy

CLEAR-EC Phase II, team PRL CorneaForge (Grand Challenge user raihan): scored model bundle and container image

These are the two files Grand Challenge scored for our Phase II submission W3R to the CLEAR-EC challenge (evaluation 6196471e-95e5-4d3b-9086-f3507abab40a, hidden-test score 8.7264). The system regresses cell density, coefficient of variation and hexagonality from one specular microscopy frame with 56 convolutional regressors in 7 groups. We publish the files for the challenge's technical review. Code, documentation and a download script that checks every file are in our review repository, shared with the organisers.

File Bytes SHA-256 Content
model_w3r.tar.gz 4,148,444,690 f7095994d5ad4f7fa15db33df5d867091ff8669814bbaea8d5f3a4ae0a57d378 the model bundle: 56 checkpoints (fp16), weights.json, calibration.json; Grand Challenge model version e98e8867-bd34-4a9e-8cc4-26036f8a4595
clear_ec_algorithm_v3_2026-11-10.tar.gz 3,652,054,132 d96f28b2203968214c61ac192d0456c0c111ffa3a53317d318ab6f0ca02c24d5 the container image (docker load names it clear_ec_algorithm_v3:latest); Grand Challenge image version 268e749f-ac08-40aa-b110-f7785ebdcf2c

The image holds no weights. It reads one frame from /input/images/corneal-specular-microscopy-image/, loads the bundle from /opt/ml/model/, and writes cell-density.json, coefficient-of-variation.json and hexagonality.json to /output/. Without the repository:

curl -L -o model_w3r.tar.gz https://huggingface.co/RINTHIRA/clear-ec-phase2/resolve/main/model_w3r.tar.gz
sha256sum model_w3r.tar.gz
mkdir model && tar -xzf model_w3r.tar.gz -C model
curl -L -o image.tar.gz https://huggingface.co/RINTHIRA/clear-ec-phase2/resolve/main/clear_ec_algorithm_v3_2026-11-10.tar.gz
docker load -i image.tar.gz
mkdir -p output && chmod a+rwx output     # input/images/corneal-specular-microscopy-image/ holds one .mha frame
docker run --rm --network none -v "$PWD/input:/input:ro" -v "$PWD/output:/output" \
  -v "$PWD/model:/opt/ml/model:ro" clear_ec_algorithm_v3

Terms: the models were trained on the CLEAR-EC training data, which is covered by the CLEAR-EC Restricted Research Data License, and three of the seven groups start from ConvNeXt-V2 weights licensed CC-BY-NC-4.0. We share the files for the challenge's technical review and for non-commercial research.