Out-of-Domain (OOD) Detection¶

FRD can detect whether newly acquired medical images come from the same domain as a reference set. This is useful for flagging potential distribution shifts — for example, images acquired with a different scanner or protocol.

Image-level OOD detection¶

Per-image scoring computes the L2 distance of each test image's radiomic feature vector from the reference set mean, then uses a statistical threshold to classify each image as in-domain or out-of-domain.

CLI¶

Bash

python -m frd_score ood path/to/reference path/to/test_images

Python API¶

Python

from frd_score import compute_frd, detect_ood

# First, extract normalised features
# (detect_ood expects pre-normalised feature arrays)
# Typically you'd get these from compute_statistics_of_paths or compute_features

results = detect_ood(
    feature_list=[ref_features, test_features],
    detection_type="image",
    seed=42,
    filenames=test_filenames,
)

print(f"Threshold: {results['threshold']:.4f}")
print(f"OOD detected: {results['predictions'].sum()}/{len(results['predictions'])}")

Output¶

Results are saved to outputs/ood_predictions/ood_predictions.csv:

Column	Description
`filename`	Image filename (or index)
`ood_score`	L2 distance from reference mean
`ood_prediction`	`True` = OOD, `False` = in-domain
`p_value`	Probability of being in-domain

Dataset-level OOD detection (nFRD)¶

The dataset-level score estimates the fraction of an entire dataset that is out-of-domain, using the area under the ROC curve.

CLI¶

Bash

python -m frd_score ood path/to/reference path/to/test_images --detection_type dataset

Python API¶

Python

results = detect_ood(
    feature_list=[ref_features, test_features],
    detection_type="dataset",
)

print(f"nFRD: {results['nfrd']:.4f}")

The nfrd score ranges from 0 (completely in-domain) to 1 (completely out-of-domain).

Configuration¶

Statistical assumptions¶

The --id_dist_assumption flag controls how the threshold is estimated from the reference scores:

Value	Description
`gaussian` (default)	Assume in-distribution scores follow a Gaussian; threshold at 95th percentile
`t`	Use Student's t-distribution with DoF = n−1
`counting`	Non-parametric; threshold at empirical 95th percentile

Validation split¶

By default, the full reference set is used for both the mean embedding and threshold estimation. To hold out a validation subset:

Bash

python -m frd_score ood path/to/ref path/to/test \
    --use_val_set \
    --val_frac 0.2

This uses 20% of the reference images for threshold estimation and the remaining 80% for computing the reference mean.

Reproducibility¶

Use --seed for deterministic results:

Bash

python -m frd_score ood path/to/ref path/to/test --seed 42