Interactive Feature Selection for Kidney Radiomics

Compare LASSO, Ridge, Elastic Net and stability selection on correlated size and shape features, and watch the selected subset move.

feature selection
LASSO
regularization
stability selection
radiomics

Features selected

2.000

Size-group selected

2.000

of 4

Shape-group selected

0.000

of 2

Coefficient magnitudes

Selected features

Feature groups are highlighted separately because size and shape features respond very differently to penalization.
Feature	Group	Coefficient / probability	Selected
tumor_volume_cm3	Size	0.074	yes
tumor_surface_cm2	Size	0.000	no
max_bbox_extent_mm	Size	0.000	no
max_axial_area_mm2	Size	0.524	yes
tumor_sphericity	Shape	0.000	no
bbox_elongation	Shape	0.000	no

Selection is itself a random variable

Ridge shrinks every coefficient but selects nothing; LASSO and Elastic Net threshold weak features to exactly zero, and the identity of the zeroed set changes with lambda. Stability selection reports how often a bootstrap replicate would have kept each feature — a single LASSO fit can be a coin flip away from a different subset.

Preparing the simulation…