Interpretability

FRD provides built-in tools for understanding what radiomic features drive the difference between two image distributions.

Overview

The interpretability analysis produces:

1. Per-feature difference ranking — which radiomic features changed the most between the two distributions (sorted by squared mean difference).
2. t-SNE visualisation — 2D embedding of the feature vectors, coloured by distribution.
3. Sorted feature difference plot — log-scale visualisation of all feature differences.

For paired datasets (same filenames in both folders), additional analyses identify which individual images changed the most.

CLI usage

Bash

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

Output is written to outputs/interpretability_visualizations/ by default. Change the output directory with --interpret_dir:

Bash

python -m frd_score path/to/dataset_A path/to/dataset_B \
    --interpret \
    --interpret_dir results/my_analysis/

Python API

Python

from frd_score import compute_frd, interpret_frd

# Option 1: Integrated with compute_frd
frd_value = compute_frd(
    ["path/to/dataset_A", "path/to/dataset_B"],
    interpret=True,
    interpret_dir="results/interp/",
)

# Option 2: Standalone (with pre-extracted features)
results = interpret_frd(
    feature_list=[features_A, features_B],  # np.ndarray each
    feature_names=feature_names,            # list of str
    viz_dir="results/interp/",
    run_tsne=True,
)

print(results["top_changed_features"][:5])

Return value

interpret_frd() returns a dict:

Key	Type	Description
top_changed_features	list[tuple[str, float]]	Top-k features by squared mean difference
n_features	int	Total number of features analysed

Dependencies

Interpretability requires optional packages:

Bash

pip install matplotlib scikit-learn

If matplotlib is missing, interpret_frd() raises an ImportError. If scikit-learn is missing, t-SNE is skipped with a warning.