Population L2-attribute model for soma identification

Statistics learned from a KC-based training set, used by aedes_soma_position() (when the cascade falls through to "l2+mesh" / "l2" / "mesh") to score how soma-like each L2 chunk of a neuron is.

Usage

aedes_soma_l2_stats

Format

A list with elements:

feature_names

Character vector of the 5 shape features used in the Mahalanobis distance.

feature_transforms

Named list documenting the raw -> feature transforms (mostly log1p).

positive_mean, positive_cov

Mean vector and covariance matrix of the positive (KC) training set in feature space.

positive_quantiles

Per-feature P5..P95 quantiles for diagnostics.

dist_npil_density

List with x (um grid), y (mixed density), y_max, support, uniform_density, uniform_eps, n – the KDE prior on signed neuropil distance.

dist_penalty_weight

Multiplier on the distance-penalty term.

n_positive, n_negative

Training set sizes.

training_set

Provenance of positives / negatives / distance prior.

build_date

Date the dataset was built.

Details

The score combines two terms:

shape

Squared Mahalanobis distance of the L2 chunk's five shape features (log_area_nm2, log_size_nm3, log_max_dt_nm, log_mean_dt_nm, roundness = pca_val_2 / pca_val_0) versus the KC positive population (positive_mean, positive_cov). Smaller = closer to a typical KC soma.

dist_npil penalty

-2 * log(f_hat(d_um) / max(f_hat)) where f_hat is the empirical KDE of signed neuropil distance over the full flywire_nuclei() population, mixed with a small uniform background so no value is hard-rejected (real central-brain neurons sit 0-3 um inside the neuropil). Weight controlled by dist_penalty_weight.

Combined as soma_score = mahal_shape + dist_penalty_weight * dist_penalty (lower = more soma-like).

See also

aedes_soma_position(); data-raw/aedes_soma_l2_stats.R for the build script.