Whole / Each Habitat Radiomics

whole_habitat

Output

whole_habitat_radiomics.csv

Definition

PyRadiomics run on the multi-label habitat map itself: the habitat label image is passed as both image and mask to PyRadiomics, so voxel intensities are habitat label IDs (not original MR/PET values). Parameters: params_file_of_habitat (optional; bundled habitat preset → habit/resources/radiomics/parameter_habitat.yaml when omitted).

Feature definitions follow PyRadiomics Feature Reference and the IBSI Phase 1 digital-phantom table on Traditional Radiomics. Default settings match PyRadiomics execute() (see PyRadiomics alignment on that page).

Output columns

Column pattern	Description
{pyradiomics_feature}	PyRadiomics feature names (no modality suffix)
(excluded)	Columns whose names contain diagnostic are dropped before export

Implementation

habit/compat/engines/habitat_extraction/habitat_features/builtin_plugins.py (WholeHabitatPlugin) → habitat_radiomics.py

each_habitat

Output

• habitat_{k}_radiomics.csv — one file per habitat index k = 1 … K (K = n_habitats from config or habitats.csv)

• habitat_count.csv — binary flags has_habitat_1 … has_habitat_K

Definition

For each habitat index k, PyRadiomics is run on the original preprocessed image with the multi-label habitat map as mask and label=k (voxels outside label k are excluded). Uses params_file_of_non_habitat (roi preset), not params_file_of_habitat. Files are written for every k in 1 … K even when a subject’s map lacks that label (empty ROI → NaN / error handling per subject).

Feature definitions follow PyRadiomics Feature Reference and the IBSI Phase 1 digital-phantom table on Traditional Radiomics. Default settings match PyRadiomics execute() (see PyRadiomics alignment on that page).

Output columns

habitat_{k}_radiomics.csv:

Column pattern	Description
{pyradiomics_feature}_of_{modality}	One column per PyRadiomics feature × image modality
(excluded)	Columns whose names contain diagnostic are dropped before export

habitat_count.csv:

Column	Description
has_habitat_{k}	1 if subject map contains label k, else 0

Implementation

habit/compat/engines/habitat_extraction/habitat_features/builtin_plugins.py (EachHabitatPlugin) → habitat_radiomics.py. v1 EachHabitatRadiomicsFeatures uses one union-bbox crop and the native C multi-label path per modality; binWidth stays per habitat (union_bin=False).

Compare habitats (cohort or one subject)

The wide each_habitat table is one row per subject. To argue that habitats are distinct – the figure a reviewer expects – melt it and contrast habitats across the cohort (paired Cliff’s delta + BH-FDR). The same objects also describe one subject (differences without p-values).

Tens-to-hundreds of texture features: draw a habitat x feature heatmap and the default features x pair Cliff’s \(\delta\) heatmap (top-k by max \(|\delta|\) when the bank is large). Pass pair=(a, b) for the single-pair lollipop. If that heatmap would be too tall, plot_habitat_feature_components() shows the same contrast on a few CVA/PCA component scores (not a 2-D embedding). Violins / bars are for a shortlist, not the full bank. Bars are one panel per feature (independent y-axis) so Energy and volume_fraction are not forced onto one linear scale.

Runnable gallery (synthetic stand-in table; swap table for your each_habitat extract):

python docs/source/examples/scripts/habitat_feature_compare_demo.py

from habit.habitat_features import compare_habitat_features, to_habitat_feature_panel
from habit.viz import (
    plot_habitat_feature_bars,
    plot_habitat_feature_components,
    plot_habitat_feature_effect,
    plot_habitat_feature_heatmap,
    plot_habitat_feature_violin,
)

panel = to_habitat_feature_panel(table)          # wide each_habitat FeatureTable
cmp = compare_habitat_features(panel)            # cohort if table has >= 2 subjects

fig = plot_habitat_feature_heatmap(cmp)          # overview (z-scored)
fig = plot_habitat_feature_effect(cmp)           # features x pair Cliff's delta
fig = plot_habitat_feature_effect(cmp, pair=(1, 2), top_k=20)  # one pair
fig = plot_habitat_feature_components(cmp)       # CVA contrast when the heatmap is too tall
fig = plot_habitat_feature_violin(cmp, max_features=6)
fig = plot_habitat_feature_heatmap(cmp, subject_id="subj001")  # one subject
fig = plot_habitat_feature_bars(cmp, subject_id="subj001", max_features=6)

compare_habitat_features(..., subject_id=...) restricts the panel first. Pairwise p_value / q_value are NaN when only one subject remains.

Runnable gallery (synthetic stand-in table; swap table for your extract): Habitat feature contrast. Graph topology columns (single_h* / pair_h*_h*) can share the joined table but do not melt through this API.

Cohort mean heatmap, z-scored per feature (plot_habitat_feature_heatmap()).

One-subject profile (plot_habitat_feature_heatmap()).

Default features x pair Cliff’s delta (star = BH q < 0.05) (plot_habitat_feature_effect()). With dozens of features the heatmap keeps the top-k by max \(|\delta|\). If that heatmap would be too tall, compare habitats on a few CVA/PCA components instead.

Same contrast, shown on 2 CVA components (use this when there are dozens of features) (plot_habitat_feature_components()).

Per-feature distributions; box + strip when n < 5 (plot_habitat_feature_violin()).

One panel per feature, independent y-axis (plot_habitat_feature_bars()).