Seeds: the CORR and PNR images

Source: minicnmfe/preprocess.py. These two summary images are how CNMF-E finds candidate neuron locations. In the main pipeline they are computed inside the greedy initialization loop (see Initialization); the standalone correlation_pnr function documented here is the same computation exposed for diagnostics.

Center-surround spatial filter

1-photon background is broad and low-frequency; neurons are compact blobs of radius ≈ sigma pixels. To suppress the former and keep the latter, frames are convolved with a center-surround PSF (make_center_surround_psf):

• a 2-D Gaussian of width sigma,
• restricted to a disk of radius 3·sigma,
• with its mean subtracted on the support, so the kernel sums to ≈ 0.

A sum-zero kernel removes any spatially uniform offset while passing neuron-scale structure. (Set center_psf=False for a plain Gaussian, or sigma=None for no filtering — not recommended for 1p.)

Per-pixel noise (PNR denominator)

estimate_noise estimates each pixel's noise standard deviation from the high-frequency power spectrum along time. It takes the real FFT of every pixel's trace, keeps the frequency band noise_range · Nyquist (default 0.25–0.5, i.e. the top half of the spectrum where calcium signal is absent), and reduces the one-sided PSD across those bins. The default reduction is logmexp — the exponential of the mean log-PSD (a geometric mean, robust to outliers) — with mean / median also available. Returns sn(H, W).

Note: estimate_noise loads the whole movie into RAM for the FFT (it needs all T frames); the package's streaming, low-RAM paths live in background.py / pipeline.py, not here.

PNR — peak-to-noise ratio

After center-subtracting the filtered movie in time,

PNR(h, w) = max_t filtered[t, h, w] / sn(h, w)

(correlation_pnr). A pixel scores high when its brightest transient is large relative to its own noise floor. PNR is a peak statistic, so it must see the full time axis — striding the max under-estimates it.

CORR — local correlation

local_correlations_fft gives each pixel the mean Pearson correlation of its trace with its 8 spatial neighbours over time. The implementation self-recenters (subtracts each pixel's time-mean) and divides by std, then averages the neighbour products via interior-slice multiplies (no FFT, despite the name). Edge pixels are divided by their actual neighbour count (5 at corners, 8 in the bulk). Before correlating, the filtered movie is thresholded at 3·sn so only suprathreshold activity contributes. Result is bounded in [-1, 1].

A real neuron lights up its neighbours together → high CORR; isolated noise does not.

The seed map

Seeds are the local maxima of CORR × PNR, keeping only pixels with CORR ≥ min_corr and PNR ≥ min_pnr (detect_seeds uses skimage.feature.peak_local_max, sorted by score descending). High on both axes means "structured and bright" — the signature of a soma. These thresholds (min_corr, min_pnr) are the primary knobs controlling how many neurons are detected.

stride subsamples time before computing the images: the per-pixel reductions tolerate it, roughly dividing wall time, with little effect on the seed map (default 1 = no subsampling).