CLEAR-EC: Corneal Learning for Endothelial Assessment and Review using AI for Endothelial Count¶

Building AI models that assess corneal graft quality directly from specular microscopy images: to make graft assessment faster, more standardized, and more scalable across eye banks.

Every corneal transplant depends on a graft whose endothelium is healthy enough to keep the cornea clear. Today, eye banks judge that quality by hand: a technician places a specular microscopy image under a device such as the KONAN, marks cells inside a small region, and reads out a handful of numbers. The process is slow, operator-dependent, and limited to a fraction of the tissue. CLEAR-EC invites the community to replace that bottleneck with automated models that read the whole image and report the metrics clinicians already trust.

Participants will use a large dataset of specular microscopy images collected from our partnering eye banks to develop models that predict three key endothelial metrics:

• Cell Density (CD) — endothelial cells per square millimeter, the primary indicator of graft viability
• Coefficient of Variation (CV) — variation in cell area (polymegethism), a marker of endothelial stress
• Hexagonality (HEX) — the proportion of cells with regular hexagonal shape (pleomorphism), reflecting endothelial health

The top-performing model will ultimately be piloted in eye banks, with the goal of making graft assessment faster, more standardized, and more scalable.

📄 More information: bit.ly/CLEAR-EC

📅 Timeline¶

All deadlines are end-of-day. Please check bit.ly/CLEAR-EC for the most current schedule and any updates.

📂 Dataset¶

The training data is publicly available on Zenodo:

➡️ https://zenodo.org/records/21270595¶

The images are specular microscopy scans of the corneal endothelium collected from our partnering eye banks. The public training set is released with ground-truth quality metrics — cell density, coefficient of variation, and hexagonality — for method development, training, and exploratory research.

The test set is hidden. It is hosted securely on Grand Challenge and used exclusively for official evaluation and ranking.

🎯 The Task¶

Given a corneal endothelial specular microscopy image, output the graft-quality metrics: cell density (CD), coefficient of variation (CV), and hexagonality (HEX).

Submissions are packaged as containerized algorithms and run on the Grand Challenge platform against the hidden test set. Each image is processed independently; your algorithm reads one specular microscopy image and writes the three predicted metrics.

🚀 Getting Started¶

1. Register for the challenge.
2. Download the training data from Zenodo.
3. Study the baseline from Github. We provide an open reference pipeline (image loading → cell segmentation → metric calculation) so you can see how raw images become CD/CV/HEX predictions and how to format your outputs. The baseline establishes a reference level that participating methods are expected to surpass.
4. Build and containerize your algorithm using the example algorithm template from the challenge phase pack.
5. Submit your algorithm on Grand Challenge before the Phase I deadline.

📈 Evaluation Metrics¶

Submissions are scored on how closely their predicted metrics match the ground truth across the three endpoints — Cell Density (CD), Coefficient of Variation (CV), and Hexagonality (HEX).

Predictions and ground truth are matched by image ID. For each metric m ∈ {CD, CV, HEX}, the absolute percent error is computed per image and averaged across the test set:

Error_m (%)   = |pred_m − gt_m| / gt_m × 100        # per image
mean_Error_m  = mean of Error_m over all images     # per metric
Overall Error = mean(mean_Error_CD, mean_Error_CV, mean_Error_HEX)

The Overall Error — the equal-weight mean of the three per-metric errors — is the ranking score. Lower is better.

CLEAR-EC — Corneal Learning for Endothelial Assessment and Review using AI for Endothelial Count.