NEWS 1/9/26: 🎉 the FRD paper has been accepted to Medical Image Analysis 🎉!

FRD (Fréchet Radiomic Distance): A Metric Designed for Medical Image Distribution Comparison in the Age of Deep Learning

By Nicholas Konz*, Richard Osuala*, (* = equal contribution), Preeti Verma, Yuwen Chen, Hanxue Gu, Haoyu Dong, Yaqian Chen, Andrew Marshall, Lidia Garrucho, Kaisar Kushibar, Daniel M. Lang, Gene S. Kim, Lars J. Grimm, John M. Lewin, James S. Duncan, Julia A. Schnabel, Oliver Diaz, Karim Lekadir and Maciej A. Mazurowski.

Fréchet Radiomics Distance (FRD)

Project Website · Paper (Medical Image Analysis) · arXiv · Evaluation Framework · Documentation · API

FRD measures the similarity of radiomic image features between two datasets by computing the Fréchet distance between Gaussians fitted to the extracted and normalized features. The lower the FRD, the more similar the two datasets.

FRD supports both 2D (PNG, JPG, TIFF, BMP) and 3D (NIfTI .nii.gz) radiological images.

Why FRD over FID, KID, CMMD, etc.?

FRD uses standardised radiomic features rather than pretrained deep features (as in FID, KID, CMMD). We show in our paper that this yields:

1. Better alignment with downstream task performance (e.g. segmentation).
2. Improved stability and computational efficiency for small-to-moderately-sized datasets.
3. Improved interpretability, because radiomic features are clearly defined and widely used in medical imaging.

FRD Versions

Installation

From PyPI (recommended)

pip install frd-score

Note: frd-score requires pyradiomics, which must be installed separately from GitHub because the PyPI release is broken for Python ≥ 3.10 (#903):

pip install git+https://github.com/AIM-Harvard/pyradiomics.git@master

From Conda (conda-forge)

conda install -c conda-forge frd-score

The conda-forge package includes all dependencies (including pyradiomics), so no additional installation steps are needed.

From source

git clone https://github.com/RichardObi/frd-score.git
cd frd-score
pip install git+https://github.com/AIM-Harvard/pyradiomics.git@master
pip install -e ".[dev]"

Requirements

• Python ≥ 3.10
• pyradiomics (installed from GitHub, see above)
• numpy, scipy, Pillow, SimpleITK, opencv-contrib-python-headless

Quick Start

CLI

# Compute FRD between two image folders (default: v1)
python -m frd_score path/to/dataset_A path/to/dataset_B

# Use FRDv0 instead
python -m frd_score path/to/dataset_A path/to/dataset_B --frd_version v0

# With masks
python -m frd_score path/to/dataset_A path/to/dataset_B -m path/to/masks_A path/to/masks_B

# Save precomputed statistics to .npz
python -m frd_score --save_stats path/to/dataset path/to/output.npz

# Re-use .npz file
python -m frd_score path/to/output.npz path/to/dataset_B

Python API

from frd_score import compute_frd

# Basic usage
frd_value = compute_frd(["path/to/dataset_A", "path/to/dataset_B"])

# With masks and options
frd_value = compute_frd(
    ["path/to/dataset_A", "path/to/dataset_B"],
    paths_masks=["path/to/masks_A", "path/to/masks_B"],
    frd_version="v1",
    verbose=True,
)

# From file lists
frd_value = compute_frd([
    ["img1.png", "img2.png", "img3.png"],
    ["img4.png", "img5.png", "img6.png"],
])

CLI Reference

Main command

python -m frd_score path1 path2 [OPTIONS]

OOD subcommand

# Image-level OOD detection
python -m frd_score ood path/to/reference path/to/test

# Dataset-level nFRD
python -m frd_score ood path/to/reference path/to/test --detection_type dataset

All shared extraction flags (--frd_version, -f, -I, --norm_ref, etc.) are also available in the ood subcommand.

Python API Reference

compute_frd(paths, **kwargs)

Main entry point. See API docs for the full signature and parameter descriptions.

save_frd_stats(paths, **kwargs)

Compute and save feature statistics to a .npz file for later re-use. Accepts the same parameters as compute_frd().

interpret_frd(feature_list, feature_names, **kwargs)

Run interpretability analysis on extracted features. Produces t-SNE plots and per-feature difference rankings. Requires matplotlib and scikit-learn.

detect_ood(feature_list, **kwargs)

Out-of-distribution detection using normalized radiomics features. Supports per-image scoring (detection_type="image") and dataset-level nFRD (detection_type="dataset").

Interpretability

FRD enables interpretable comparison of image sets. Use the --interpret flag or call interpret_frd() to:

• Rank the most-changed radiomic features between two distributions
• Visualise feature distributions via t-SNE
• Identify which images changed the most (for paired datasets)

python -m frd_score path/to/dataset_A path/to/dataset_B --interpret

Out-of-Domain (OOD) Detection

FRD can detect whether newly acquired medical images come from the same domain as a reference set — useful for flagging potential distribution shifts (e.g. different scanners, protocols).

# Per-image OOD scores and p-values
python -m frd_score ood path/to/reference path/to/test_images

# Dataset-level OOD score (nFRD)
python -m frd_score ood path/to/reference path/to/test_images --detection_type dataset

Results are saved to outputs/ood_predictions/ood_predictions.csv with columns: filename, ood_score, ood_prediction, p_value.

Citation

If you use this library in your research, please cite:

@article{konz2026frd,
    title     = {Fr\'{e}chet Radiomic Distance (FRD): A Versatile Metric for
                 Comparing Medical Imaging Datasets},
    author    = {Konz, Nicholas and Osuala, Richard and Verma, Preeti and
                 Chen, Yuwen and Gu, Hanxue and Dong, Haoyu and Chen, Yaqian
                 and Marshall, Andrew and Garrucho, Lidia and Kushibar, Kaisar
                 and Lang, Daniel M. and Kim, Gene S. and Grimm, Lars J. and
                 Lewin, John M. and Duncan, James S. and Schnabel, Julia A. and
                 Diaz, Oliver and Lekadir, Karim and Mazurowski, Maciej A.},
    journal   = {Medical Image Analysis},
    volume    = {110},
    pages     = {103943},
    year      = {2026},
    publisher = {Elsevier},
    doi       = {10.1016/j.media.2026.103943},
    url       = {https://www.sciencedirect.com/science/article/pii/S1361841526000125},
}

Earlier FRD work:

@article{osuala2024towards,
    title   = {Towards Learning Contrast Kinetics with Multi-Condition
               Latent Diffusion Models},
    author  = {Osuala, Richard and Lang, Daniel and Verma, Preeti and
               Joshi, Smriti and Tsirikoglou, Apostolia and Skorupko, Grzegorz
               and Kushibar, Kaisar and Garrucho, Lidia and Pinaya, Walter HL
               and Diaz, Oliver and others},
    journal = {arXiv preprint arXiv:2403.13890},
    year    = {2024},
}

Links

• API Documentation — overview, benchmarks, datasets, FAQ
• Change Log — overview, benchmarks, datasets, FAQ
• Project Website — overview, benchmarks, datasets, FAQ
• Journal Article — Medical Image Analysis, Vol. 110 (2026)
• arXiv Preprint
• Evaluation Framework — scripts for OOD detection, translation evaluation, and metric comparison
• API Documentation — full docs hosted on GitHub Pages

Acknowledgements

• Preeti Verma — implementation of a script of an early frd version.
• Nicholas Konz — FRDv1 and RaD repository
• PyRadiomics — radiomic feature extraction backend
• pytorch-fid — Fréchet distance implementation reference

License

Apache 2.0