Triaged ML

---

About

Summary

This algorithm measures three commonly reported features of the corneal endothelium from specular microscopy images:

• Cell Density (CD): the estimated number of endothelial cells per mm²
• Coefficient of Variation (CV): variation in cell size
• Hexagonality (HEX): the proportion of cells with six neighbours

The method combines an auditable image-analysis pipeline with a robust machine-learning correction layer. Cell borders are detected through two independent segmentation routes, and measurements are calculated from the resulting cell geometry. A quality-aware ensemble then makes small, bounded corrections using image texture, segmentation quality, agreement between alternative measurements, and other reliability indicators.

The machine-learning layer uses two different experts: a median-error gradient-boosted model and a randomized-tree model. When the experts disagree, the correction is automatically reduced toward the classical result. This makes the algorithm less likely to apply an aggressive correction to an unfamiliar or poor-quality image.

The current model was developed using 5,982 labeled corneal specular microscopy images. Performance was evaluated using five-fold, slide-grouped cross-validation.

Mechanism

How the algorithm works¶

The image passes through the following stages:

1. Input and calibration

   The image is loaded and converted into a single grayscale microscopy frame. Pixel calibration is read from the image metadata when available. A validated fallback calibration is used when suitable metadata are unavailable.

2. Image preparation

   The pipeline corrects uneven illumination and creates several image representations that highlight cell borders, local contrast, and ridge-like structures.

3. Primary segmentation

   A marker-controlled watershed identifies individual endothelial cells from the enhanced border pattern.

4. Independent segmentation

   A distance-transform watershed analyzes the image through a second route. This branch is designed to behave differently from the primary segmentation and provides an independent check on the cell mosaic.

5. Segmentation repair and agreement

   The two segmentation results are compared. Conservative repairs are made only where the independent evidence supports them. Areas where the branches disagree contribute to lower reliability.

6. Quality and analysis region

   The algorithm estimates which parts of the image are sufficiently focused and visible for measurement. It selects a coherent analysis region containing gradable cells and avoids poorly defined peripheral regions when possible.

7. Classical measurements

   CD, CV, and HEX are calculated from cell area, size variation, and cell-neighbour relationships. Alternative versions of each measurement are also calculated for comparison.

8. Classical calibration

   A slide-grouped calibration model corrects systematic measurement bias while preserving the original measurement definitions.

9. Robust machine-learning correction

   Two nonlinear experts analyze the calibrated values together with:

   • alternative CD, CV, and HEX estimates;
   • segmentation agreement;
   • cell counts and reliability;
   • focus, contrast, and gradability;
   • local image texture and gradient patterns;
   • frequency-domain image information.

   CD and CV corrections are applied in logarithmic space, while HEX is corrected in logit space. This keeps the corrections relative to the original value and helps keep the outputs within meaningful ranges.

10. Safety checks

    When the two experts disagree, the correction is reduced toward the classical calibrated result. Corrections and final outputs are also restricted to plausible ranges. If the robust model cannot be loaded or evaluated, the algorithm automatically returns the classical calibrated result.
    

Inputs¶

The underlying container supports planar grayscale microscopy images stored as:

• .mha
• .mhd
• .tif or .tiff
• .png

Grand Challenge normally provides image inputs to the container as MetaImage files. A two-dimensional image represented as a singleton three-dimensional MetaImage is supported. True multi-slice volumes are not supported.

Outputs¶

The algorithm produces:

• Cell Density: cells/mm²
• Coefficient of Variation: a fractional value, where 0.40 represents 40%
• Hexagonality: a fractional value, where 0.55 represents 55%

A diagnostic confidence sidecar may also be produced. Confidence represents estimated measurement reliability and should not be interpreted as disease probability or corneal health.

---

Interfaces

This algorithm implements all of the following input-output combinations:

Validation and Performance

Internal validation¶

The robust correction model was evaluated using five-fold, slide-grouped cross-validation. Images belonging to the same slide_id group were kept together to reduce leakage between training and validation folds.

A total of 5,982 labeled images were evaluated. Performance is reported as absolute percentage error relative to the reference measurements.

Overall results¶

Metric	Classical median error	Robust median error	Classical mean error	Robust mean error	Classical 90th-percentile error	Robust 90th-percentile error
Cell Density	7.94%	6.55%	12.04%	9.96%	23.30%	19.50%
Coefficient of Variation	9.97%	9.28%	12.69%	11.74%	25.70%	23.92%
Hexagonality	9.08%	8.96%	11.52%	11.30%	23.60%	22.93%
Combined average	9.00%	8.26%	12.08%	11.00%	24.20%	22.11%

The combined median error decreased from 9.00% to 8.26%. The combined mean error decreased from 12.08% to 11.00%, and the combined 90th-percentile error decreased from 24.20% to 22.11%.

The largest improvement was seen for Cell Density. Improvements for CV and HEX were smaller but remained consistent across the overall median, mean, and 90th-percentile summaries.

Results by image quality¶

Quality groups are based on the estimated fraction of the image that is sufficiently gradable:

• Low quality: less than 10% gradable
• Medium quality: 10% to less than 50% gradable
• High quality: 50% or more gradable

Quality	Metric	Images	Robust median error	Robust mean error	Robust 90th-percentile error
Low	Cell Density	762	7.53%	11.84%	24.08%
Low	Coefficient of Variation	762	10.22%	12.00%	25.15%
Low	Hexagonality	762	9.28%	12.58%	25.57%
Medium	Cell Density	1,460	7.03%	11.07%	21.55%
Medium	Coefficient of Variation	1,460	9.59%	11.58%	23.83%
Medium	Hexagonality	1,460	9.48%	11.68%	23.98%
High	Cell Density	3,760	6.23%	9.14%	17.97%
High	Coefficient of Variation	3,760	9.06%	11.75%	23.66%
High	Hexagonality	3,758	8.70%	10.90%	22.14%

Performance is strongest on high-quality images. Low-quality images have wider error distributions and should be interpreted more cautiously.

These results represent internal development cross-validation. They are not independent, prospective, or external clinical validation.

---

Challenge Performance

Date	Challenge	Phase	Rank
Aug. 10, 2026	CLEAR-EC	Phase I: Screening Phase (Preliminary Evaluation)	7

Uses and Directions

This algorithm was developed for research purposes only. It may be used to:

• estimate Cell Density, Coefficient of Variation, and Hexagonality from corneal specular microscopy images;
• support retrospective research and dataset analysis;
• provide standardized preliminary measurements for human review;
• identify images where low quality or estimator disagreement may reduce reliability;
• compare endothelial morphology across research cohorts when image acquisition is reasonably consistent.

For best results:

1. Upload one planar corneal specular microscopy image per test case.
2. Use an image where the endothelial cell mosaic is visible and reasonably focused.
3. Confirm that the image contains corneal endothelium rather than another microscopy modality.
4. Review low-quality images and unexpected measurements manually.
5. Treat the measurements as quantitative research estimates, not as standalone clinical conclusions.

Warnings

• Research use only. This algorithm has not been established as a medical device and has not been prospectively validated for clinical decision-making.

• Image quality matters. Blur, glare, low contrast, incomplete cell borders, compression artifacts, and very small gradable regions can increase measurement error.

• Performance varies by quality. Errors are generally higher and less predictable on low-quality images.

• Device generalization is not guaranteed. Images from devices or acquisition settings that differ from the development data may produce systematic bias.

• True volumetric images are not supported. Singleton three-dimensional MetaImages that contain one planar frame are accepted, but multi-slice volumes are rejected.

• Confidence is not a diagnosis. Confidence describes estimated measurement reliability. It does not represent disease probability, endothelial health, or the likelihood that treatment is needed.

• Confidence and robust correction are separate layers. The diagnostic confidence estimate is primarily calibrated around the classical measurement pipeline. It should be used as supporting information rather than a guarantee of the final robust-corrected output.

• Automated correction has limits. The robust ensemble reduces its correction when its experts disagree and falls back to the classical result if necessary, but these safeguards cannot detect every unfamiliar image.

• Very poor segmentation may still produce an unreliable result. Images with few visible cells or no coherent endothelial mosaic should be reviewed or excluded.

• Population-level validation is limited. Performance has not been fully established across demographic groups, disease subtypes, prior corneal surgery, contact-lens conditions, or pediatric populations.

Common Error Messages

Error message	Meaning and solution
Could not find manifest.json in the container image file. Was this created with docker save?	A model archive was uploaded in the Containers section. Upload the Docker-save archive under Containers and upload the algorithm model archive separately under Models.
algorithm model with this checksum already exists	The same model archive has already been uploaded. Select or reactivate the existing model instead of uploading an identical copy.
Only planar MetaImages are supported	The uploaded file is a true multi-slice volume. Upload a single two-dimensional microscopy frame. Singleton 3-D MetaImages containing one frame are supported.
No supported input images found	The input directory did not contain a supported image. Upload the image through the algorithm's defined image input rather than uploading a ZIP archive as the test image.
Inference produced no finite cell-density result	No usable endothelial segmentation was produced. Confirm that the image contains a visible corneal endothelial mosaic and is not blank, severely blurred, or from another modality.
Invoke endpoint returned status 500, expected 201	The algorithm encountered an input or processing error. Open the result logs and look for the first error immediately before the final HTTP 500 message.
Time limit exceeded	Confirm that the newest optimized container is active, that only one image was submitted for the test case, and that the algorithm is configured for CPU execution.
WARNING: no calibration model found	The algorithm model archive is missing or is not active. Upload and activate the robust model archive under Models.
WARNING: robust correction failed; using calibrated baseline	The nonlinear correction could not be evaluated. The algorithm safely returned the classical calibrated result. Review the input and confirm that the robust model archive is active.
Container import fails with manifest.json missing	Use the container file created with docker save, not the model tarball or source-code ZIP.
Model import completes but the robust correction is not shown	Confirm that robust_residual_ensemble_v2.pkl is present in the active model archive and that the newest container version is active.