Suite2p: beyond 10,000 neurons with standard two-photon microscopy#
Some main takeaways from the publication.
Steps#
image registration
region-of-interest (ROI) detection;
ROI labelling and quality control;
activity extraction with neuropil correction and spike deconvolution.
Claims:#
Suite2p does not assume that cell activity is indepenedent of neuropil
Finds 2x as many neurons as CNMF
Preferrale to algorithms based on greedy segmentation of neighboring pixels
Registration#
Many techniques find the maximum correlative peak between reference and target.
Disadvantage that it relies on low-frequencys: neuropil, small graduents of pixel fluctuation
High frequency is essential for registration, i.e. edges, activity
ROI Detection#
Deconvolution#
Core is a model that assumes the signal of an individual pixel is the sum of signals originating in distinct active compartments (soma, dendrite, spine) and more diffuse contamination from neuropil
CNMF extraction significantly comprimised by subsampling (why?)
Neuropil#
produces contamination from average activity of out-of-focus dentrides/axons
2p scopes have PSF one magnitude wider in Z vs X/Y
So, single plane is really a weighted average of a volume extenting 10’s of micron
(a) Example cell (blue) and neuropil areas (red-shaded annuli) superimposed on the variance image (white), for an imaging session in superficial V1. The neuropil signal is defined as the signal inside the red-shaded annuli but not in any detected ROIs.
The neuropil signal is highly correlated (related) to the signal recorded from neurons (somata).
The neuropil signal varies slowly across space (it changes gradually over distance).
Sometimes, when the neuropil signal gets very large, it reflects periods of high network activity (when neurons are firing a lot).
This is modeled through spatial basis functions
Spatial Basis Functions:
Imagine dividing the whole image into small overlapping “tiles” (like tiles on a floor).
Each tile represents a smooth, predictable pattern of how the neuropil signal changes across space.
These tiles are modeled using “raised cosines” (smooth, circular patterns) that are sized based on the average size of a neuron.
ROI Signals:
Each neuron has its own unique activity pattern over time.
Each pixel in an ROI contributes to the same timecourse (signal over time) for that neuron, scaled by how strongly that pixel is part of the neuron.
Fitting the model:
The recorded signal at each pixel is the sum of three things:
The neuron activity (signals from the ROIs).
The neuropil signal (from the overlapping tiles).
Measurement noise (random errors in the recording).
Super res mean image:
brain constantly in motion, generating continuous x/y offsets
same as registration
instead of shifting frame, bin frames by their x/y offset @subpixel resolution
get mean-image for each XY bin, combine into super image
Suite2p classifier#
labels cells as cell/noncell based on statistics
activity dependent: skewness, variance, corr with surrounding px
anatomy dependent: area, aspect ratio
quality metric: posterior probability tht each ROI == soma
Extraction and Deconvolution#
Get final flourescent signal.
Skipping lots of math.
Remove neuropil contamination
(sometimes useful) estimate spike train underlying the calcium trace
Recommendation: unconstrained non-negative deconvolution using exponential kernels
Benchmark vs CNMF#
Dataset: [5000x512x512] with 64GB RAM
Single CNMF optimization: 8 minutes
Single Suite2p optimization: 7 seconds
Large difference due to difference extraction algorithm / sparse matrix
SVD Low rank approximation
CNMF can’t use this approx, invalidates the positivity constraints CNMF imposes on calcium dynamics
CNMF uses a single global neuropil signal model
Suite2p with this “single neuropil” model == lots of false ROI’s
Estimate that CaImAn does the same
Essentially, a global neuropil means that all pixels recieve the same contribution of contamination.