Auto-evaluation

Source: minicnmfe/evaluate.py (auto_evaluate_components, spatial_r_values) and pipeline.py:CNMFe.evaluate. The final extraction step records per-component quality metrics.

Report-only by default; the gate is opt-in. evaluate() always writes the per-component model.eval_info (snr_amp, pixel_count) for inspection, but the acceptance gate is off by default (auto_eval_snr_amp_thr = 0.0), so model.accepted_mask is all-True — every extracted component is accepted. The rationale: with seed thresholds (min_corr/min_pnr) set to the recording's noise floor you don't get ghosts, so a default gate mostly produces false negatives (rejecting real dim cells). Control ghosts upstream; raise auto_eval_snr_amp_thr (~3) and/or min_pixel only to opt in to filtering on a noisy recording.

Nothing is ever dropped. Even with the gate on, evaluation only writes the boolean model.accepted_mask (K,); you filter downstream yourself (model.A[:, model.accepted_mask], and likewise for C/S/YrA/C_raw). This keeps the step non-destructive and re-runnable on a loaded model to retune thresholds without re-extracting.

The two checks (when opted in)

With the gate enabled, a component is accepted only if it passes both (auto_evaluate_components):

1. Pixel-count floor — footprint has at least min_pixel non-zero pixels.

2. Mean-amplitude SNR — scale-invariant, the real discriminator:

snr_amp[k] = (‖a_k‖² / npix[k]) / mean( sn[support_k]² )

i.e. the mean squared footprint amplitude over its support, divided by the mean pixel-noise variance over the same support, thresholded at auto_eval_snr_amp_thr (default 0.0 = off; use ~3.0 when opting in). A real sigma=3 Gaussian footprint scores ~10–70; a ghost component (born from a background-noise seed under loose init thresholds) sits at or below ~2 even when it has many pixels — so a pure pixel-count filter cannot separate them, but this SNR check can. Set the threshold to 0 to accept every component on the SNR check.

The a_norm subtlety

snr_amp ∝ ‖a_k‖². But the pipeline's final step relabels footprints to unit L2 norm (amplitude moved into the traces — see the overview), which would flatten ‖a_k‖² to 1 and destroy the discriminator. To avoid that, the original norms are cached on model.A_norm and passed in as a_norm, so ‖a_k‖² is reconstructed as a_norm[k]² instead of read from the now-unit-norm A. Passing a_norm=None reads directly from A (correct only for un-normalized footprints — old saved models / the unit tests).

Spatial r-value (separate diagnostic)

spatial_r_values is a complementary spatial quality metric, not part of the default accept/reject gate. For each component it builds the activity image ΔF = mean(movie over the trace's peak frames) − mean(movie over all frames) on the footprint's bounding box plus a pad-pixel surround, and Pearson-correlates that with the footprint values. A clean "bright blob on dark surround" scores high; a merged or sprawled footprint spanning several activity spots scores low — catching shape problems the temporal SNR is blind to.