henhaobio
Sign in

bioRxivNeurosciencePreprint

Suite2p: beyond 10,000 neurons with standard two-photon microscopy

Marius PachitariuiD, Carsen StringeriD, Mario DipoppaiD, Sylvia SchröderiD, Luigi Federico RossiiD, Henry William Peter DalgleishiD, Matteo CarandiniiD, Kenneth D. Harris

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Two-photon microscopy of calcium-dependent sensors has enabled unprecedented recordings from vast populations of neurons. While the sensors and microscopes have matured over several generations of development, computational methods to process the resulting movies remain inefficient and can give results that are hard to interpret. Here we introduce Suite2p: a fast, accurate and complete pipeline that registers raw movies, detects active cells, extracts their calcium traces and infers their spike times. Suite2p runs on standard workstations, operates faster than real time, and recovers ~2 times more cells than the previous state-of-the-art method. Its low computational load allows routine detection of ~10,000 cells simultaneously with standard two-photon resonant-scanning microscopes. Recordings at this scale promise to reveal the fine structure of activity in large populations of neurons or large populations of subcellular structures such as synaptic boutons.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 2 (20 Jul 2017) · CC BY 4.0

Suite2p processes two-photon calcium movies faster than real time on desktop workstations, roughly doubling curated cell yield over CNMF and recovering 13,451 neurons in one session.

Earlier calcium-imaging pipelines were slow, memory-hungry or cluster-dependent, and constrained demixing methods such as CNMF (constrained non-negative matrix factorization) can disguise data defects while recovering fewer cells. Suite2p separates processing into inspectable stages with a spatially varying neuropil model and singular value decomposition (SVD) compression, making ~10,000-neuron recordings routine on inexpensive workstations.

Key results

  1. ShownSuite2p processed large-scale multi-plane recordings on a single workstation, recovering 13,451 neurons from one session, whereas CNMF could not run on these datasets because memory needs exceeded 64 GB even with temporal subsampling.

    2 hour, 11-plane session with 900 μm field of view; automated sorting ~10 minutes/plane (Core i7, 32 GB RAM, GTX 980Ti) plus ~10 minutes/plane curation. On subsampled standard data, one iteration took ~7 seconds vs ~8 minutes for CNMF; SVD compression gave 10-100x acceleration. · Application to single and multi-plane data and…; Accelerating cell detection with large-scale SVD…; Figure legends

  2. ShownSuite2p yielded nearly twice as many curator-confirmed somata as CNMF at matched region-of-interest (ROI) totals and about three times as many as autoROI, consistent with an ablation where a single global neuropil term (as in CNMF) diverted ROIs to neuropil.

    Four GCaMP6m datasets from mouse primary visual cortex (V1), cells counted manually if soma-shaped with transients absent from nearby neuropil; ROC on each method's quality score favored Suite2p; autoROI averaged 241 vs 792 Suite2p cells. Ablation: no or one global neuropil component left most ROIs on neuropil; 25 basis functions (default) gave almost all cellular ROIs. · Stage 2: ROI detection; Application to single and multi-plane data and…; Figure legends

  3. ShownOn the public Neurofinder challenge, whose labels come from human curators, Suite2p scored close to but below the top-ranked algorithm, HNCcorr.

    Suite2p 0.571 vs 0.597 for HNCcorr; the comparison omitted datasets 1, 2 and 8 of the 9 test sets, whose manual annotations ignored activity and on which anatomy-based algorithms do significantly better. · Discussion; Neurofinder benchmark

  4. ShownA naive Bayes classifier trained on curator decisions, using activity and shape statistics, reduced the non-somatic fraction of accepted ROIs while losing few true somata; its posterior probability doubles as a per-ROI quality score.

    Large-scale V1 recordings: raw stage-2 output held 15% non-somas per human curation; after training, 5% of auto-accepted ROIs were non-somas (true-positive rate 85% to 95%) with 1.5% of curated somata missed; stricter thresholds gave 99% true positives but almost 50% misses. · Stage 3: ROI labelling and quality control; Automatically labelling ROIs as somata with…; Figure legends

  5. ShownSubpixel phase correlation, which whitens images and interpolates the correlation peak by kriging, registered simulated movies more accurately than cross-correlation variants at the speed of non-subpixel cross-correlation.

    Simulations with known rigid shifts and realistic noise (errors averaged over 10 randomized runs): residual error ~0.1 pixels across eight varied parameters while standard methods lagged; for 512×512 frames, >15 times faster than 10x upsampled cross-correlation, plus ~4 times further speedup on GPU. · Stage 1: image registration via phase…; Figure legends

Methods, data and reproducibility

Four separable stages: subpixel phase-correlation registration (rigid or block non-rigid); ROI detection fitting localized sources plus smooth raised-cosine neuropil bases to SVD-compressed movies; naive Bayes cell labelling with GUI curation; trace extraction jointly estimating per-cell neuropil coefficients and non-negative spike deconvolution.

  • Standard-size awake mouse V1 recordings after viral GCaMP6m expression, 512×512 pixels over ~500μm at 5Hz, used for CNMF and autoROI yield benchmarks · mouse · n = four datasets
  • Large-scale awake V1 recordings from transgenic mice with GCaMP6s in all excitatory neurons, 11 planes, ~900μm field of view, >1,000 cells per plane · mouse, GCaMP6s transgenic (Ai94; Camk2a-tTA; Emx1-Cre in Figure 6) · n = two datasets; one 2 hour session gave 13,451 cells
  • 2 hour, 10-plane recording over 900μm x 900μm sampled at 3 Hz per plane, used to report per-stage processing time · mouse · n = 9623 neurons
  • Nine-plane layer 2/3 V1 recording with 112 natural-image stimuli, used for spike-deconvolved population rasters · mouse (Camk2a-tTA; Ai94(TITL-GCaMP6s); Rasgrf-Cre) · n = 7,167 neurons
  • High-magnification recordings of GCaMP6s-labelled thalamic (lateral geniculate nucleus) axonal boutons in V1 and of sparsely labelled layer 1 dendrites and spines · mouse · n = two simultaneously imaged fields of view each (axons; dendrites)
  • Simulated movies with known rigid shifts, realistic noise and activity-dependent fluorescence, plus line-by-line non-rigid motion, for registration validation · n = averaged over 10 randomized simulations
  • Neurofinder challenge test datasets from several laboratories, annotated by human curators using differing methods; datasets 1, 2 and 8 excluded from the HNCcorr comparison · n = 9 test datasets
  • Same fields of view imaged on day 1 and day 31, used to validate affine cross-session ROI matching · mouse · n = 2 fields of view across 2 mice
Methods & tools
Suite2p (this work: registration, ROI detection, curation GUI, deconvolution), Phase correlation with kriging subpixel interpolation, Block-wise non-rigid registration with Gaussian basis interpolation, Approximate SVD low-rank compression, Naive Bayes ROI classifier, CNMF (constrained non-negative matrix factorization) baseline, OASIS spike deconvolution (wrapped in Suite2p without L1 penalty), autoROI greedy pixel-segmentation baseline, HNCcorr (top-ranked Neurofinder algorithm), Neurofinder benchmark, GCaMP6m and GCaMP6s calcium indicators, Thorlabs resonant-scanning two-photon microscope
Reagents & environment
Mice: wild-type, Ai94;CamK2a-tTA;Emx1-Cre, Camk2a-tTA;Ai94;Rasgrf-Cre and Gad-Cre lines; AAV2/1 CamKII-Cre, hSyn-flex-GCaMP6s, CAG-flex-tdTomato and hSyn-GCaMP6m/s vectors (University of Pennsylvania Viral Vector Core); no software versions stated.
Assay conditions & compute
Tested on Intel i7 PCs with 16,32 or 64 GB of RAM, SSDs and GTX 970/980Ti/1060/1070/1080 GPUs ($1000-$2,000); 13,451-cell session sorted at ~10 minutes/plane (i7, 32 GB RAM, GTX 980Ti); imaging on a Thorlabs resonant-scanning microscope at 2.5Hz.
Reproducibility gaps
Imaging datasets have no stated repository or accession (only a YouTube clip of one large-scale recording is linked); no software versions are given; baseline CNMF and OASIS repository URLs are printed with embedded spaces.
Limitations (6)
  • CNMF could run only with temporal subsampling (up to 5,000 frames at 512×512 on a 64GB RAM workstation), which severely degraded its activity extraction, so the comparison covers cell detection only, not trace quality. · stated by the authors
  • Neurofinder ground truth comes from human curators, who may miss rarely firing cells in short (5-10 minute) recordings, so the authors caution that benchmark scores could penalize correct Suite2p detections. · stated by the authors
  • Large-scale yields also depended on acquisition choices (slow GCaMP6s for high signal-to-noise ratio, 3Hz scan rates, laser power near 100mW), so >10,000-cell counts are not attributable to Suite2p alone. · stated by the authors
  • Automation is incomplete: rigid registration can leave residual non-rigid motion that users must spot by visual inspection, and the trained classifier still misses 1.5% of curated somata, with misses reaching almost 50% at thresholds giving 99% true positives. · stated by the authors
  • Yield comparisons rest on manual soma counts by the developers at matched ROI totals, with no stated blinding or independent ground truth such as electrophysiology, so curator judgment underlies the reported yield advantage. · noted in this summary
  • Neuropil-coefficient fitting and the default unconstrained deconvolution are justified by method comparisons and ground-truth spiking data that are referenced but not presented here, so trace and spike accuracy remain unvalidated in this manuscript. · noted in this summary

Read this if you process two-photon calcium imaging from thousands of neurons or from boutons and spines and want a workstation-scale pipeline with explicit neuropil modeling and fast manual curation.

Citations
1,254

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISStringer '26Hoang '20Berens '18Clough '21Ledochowits '19Hoang '17
78% semantic matchNeuroscience
Carsen Stringer, Chris S Ki, Nicholas A. Del Grosso+3 more·bioRxiv·Feb 6, 2026· 9 citations