CT-eval / README.md
Kiwi321's picture
Add files using upload-large-folder tool
1c87a9c verified

CT Generation Evaluation Docker

This Docker container evaluates CT volume generation predictions by running multiple metrics and merging their outputs into a single JSON file.

Metrics

  • FVD_CTNet – Frechet Video Distance computed on 3D CT volumes using the CT-Net backbone.
  • CLIPScore / CLIP_I2I – CLIP-based image-text similarity, reporting both I2T and I2I scores and their mean.
  • FID_2p5D – 2.5D Frechet Inception Distance computed on orthogonal slices (XY, XZ, YZ).

Input Specification

The container accepts either:

# Mount predictions directory or ZIP archive to /input
docker run --rm \
  -v "$(pwd)/input":/input \
  -v "$(pwd)/output":/output \
  forithmus/ctgen-eval:latest

Inside /input, provide either:

  • A set of flattened .mha files under /input or any nested subdirectories.
  • A single .zip archive anywhere under /input containing .mha files.

Notes:

  • The first matching .mha files or the first .zip found will be used for evaluation.

Ground-Truth Data

Ground-truth .mha volumes are baked into the container at:

/opt/app/ground-truth

Each file in this directory should be named by accession or unique identifier and have a .mha extension.

Output Specification

After evaluation, the container writes the merged metrics JSON to /output/metrics.json. The file has the following structure:

{
  "FVD_CTNet": <float>,        // FVD score
  "CLIPScore": <float>,        // CLIP I2T score
  "CLIPScore_I2I": <float>,    // CLIP I2I score
  "CLIPScore_mean": <float>,   // mean CLIP score
  "FID_2p5D_Avg": <float>,     // average 2.5D FID
  "FID_2p5D_XY": <float>,      // XY-plane FID
  "FID_2p5D_XZ": <float>,      // XZ-plane FID
  "FID_2p5D_YZ": <float>       // YZ-plane FID
}

All values are floats rounded to four decimal places.

Testing

To verify functionality, run:

./test.sh

Ensure the script has execute permissions:

chmod +x test.sh

Exporting

Use the export.sh script to set environment variables before running evaluation:

source ./export.sh

This will generate a .tar.gz file for submission to the challenge platform.

For questions or issues, please contact the challenge organizers.