10  Derived Features

FLIM Playground lets users compose new derived features by combining existing features, potentially from different channels. An example would be the normalized optical redox ratio, composed from the intensity_sum feature from each of the two channels: NAD(P)H / (NAD(P)H + FAD).

10.1 Build Derived Features

The builder is revealed when the Derived features checkbox near the bottom of the Configuration page is ticked.

Any derived features already defined for the current profile are listed at the top, each showing its name, its expanded formula, and a 🗑️ button to delete it. Below the list, a new feature is composed from:

  • a Template (a common formula, or Custom… to write your own),
  • a Name (e.g. redox_ratio) — the output column becomes Derived: <Name>, and
  • the operands that fill the formula.

The operands/features available for selection should span all channels. If no channel has a feature extractor selected yet, the builder shows “Select feature extractors for your channels first — there are no features to derive from yet.” — configure the extractors first, then come back.

Operands come only from the categorized numerical features. Raw bookkeeping columns — such as amp1, offset, reduced_chi_square, the centroid coordinates, and the a2 / a3 fit remainders — are intentionally excluded from the operand list and cannot be used as derived-feature inputs.

Derived features are stored per profile, so each extraction profile keeps its own set.

10.1.1 Templates

Choosing a Template other than Custom… fills in the formula for you and renders one searchable dropdown per operand slot (A, B, …). Four ready-made templates are provided:

Each takes two operands: pick a feature for A and a feature for B, give the result a Name, and the live preview confirms it. The fifth option, Custom…, unlocks the free formula editor described next.

10.1.2 Custom Formulas

Custom… mode lets you compose an arbitrary formula. First pick the operands from the Operands (each becomes A, B, C… in the order picked) multiselect — the features you choose are aliased A, B, C, … in pick order, and a caption echoes the mapping (e.g. A = Intensity texture_fad: intensity_sum, B = …). Then build the formula in the Formula box, typing directly or using the append buttons for the operators (+ - * / ( )) and for each alias.

The formula language supports:

  • the four arithmetic operators +, -, *, /,
  • a leading sign (unary + / -),
  • numeric constants, and
  • parentheses for grouping.

Division by zero yields NaN rather than raising an error or producing infinity. Anything outside this language — function calls, exponentiation (**), modulo (%), comparisons, or a bare operand with no operator — is rejected.

10.2 Compute Derived Features

Derived features are computed during Numerical Feature Extraction, right after each field of view’s base features are extracted and concatenated. The definitions are baked into the FOV metadata, so re-running extraction from a saved metadata file reproduces the same derived columns. Each definition adds one Derived: <name> column to the output dataset.

If any cell ends up with a NaN in a derived column (for instance from a divide-by-zero or a NaN operand), a single dataset-wide warning is shown for that feature — “Derived feature name has N NaN value(s) out of M cell(s) (e.g. from divide-by-zero or a NaN operand).” As with other features, cells with NaN are not dropped from the output.

10.3 Derived Feature In Data Analysis

On the Data Analysis page, all derived features are collected into one Derived Features group in the numerical selection widgets, alongside the per channel-extractor groups. A derived feature is labeled by its plain name (e.g. redox_ratio); unlike extracted lifetime or intensity features, it carries no unit or symbol, since its meaning is entirely user-defined.