By: darioringach
M: Analysis, Uncategorized, , , , , , , ,

An extremely simple deconvolution algorithm

A number of sophisticated deconvolution algorithms have been proposed in the literature, most of which involve solving an L_1 optimization problem.

While trying to develop a fast, on-line spike detection algorithm for Scanbox I tried a simple idea that seems to work reasonably well and the resulting algorithm is trivial.

In all likelihood this must have been published somewhere.  If you happen to know a relevant reference, just let me know in the comments below.

Suppose the calcium signal, y_n, is the response of a linear system with an exponential impulse response to a train of impulses u_n.

Then, the response satisfies y_n = \alpha y_{n-1} + u_n + \epsilon_n, where we assume \epsilon_n is Gaussian noise.

One simple way to estimate u_n is to first find the value of \alpha that minimizes the error  E \{ ( y_n - \alpha y_{n-1} - u_n )^2 \}.

A few calculations yield \alpha = ( \mu^2 - m_{12}  ) /   ( \mu^2 - m_{02}  ), where \mu = E\{ y_n \}, m_{12} = E\{y_n y_{n-1}\} and m_{02} = E\{y_n^2\}.

As all these quantities are easily estimated recursively as new data arrive, one can also maintain a recursive estimate of the optimal \alpha.

Then, the estimated input to the system can be estimated by computing \hat{u}_n = y_n -  \alpha y_{n-1}.  So, all we need to do is to keep a recursive estimate of \alpha and the last two samples of the signal.

Below is an example of how this simple algorithm performs.deconv_easyA simulated spike train, u_n (second panel form top) is convolved with an exponential impulse response and corrupted by additive Gaussian white noise to yield a simulated signal y_n (top panel).  The third panel shows the estimated \hat{u}_n. Finally, we threshold this signal to obtain the estimated spike train shown in the bottom panel (thresholding is done by means of Otsu’s algorithm, implemented by Matlab’s gray thresh).

Here is the code if you want to play with it:

% generate ca-trace

u = double(rand(1,10000)<0.1); % original spikes
h = exp(-(0:40)/10);       % exponential impulse response
y = conv(h,u);             % convolve
y = y+0.15*randn(size(y)); % additive gaussian noise 

% estimate alpha, uhat and input spike train
m = mean(y);
m02 = mean(y.^2);
m12 = mean(y(2:end-1) .* y(1:end-2));
a = (m^2-m12) / (m^2-m02);
uhat = [0 y(2:end) - a * y(1:end-1)];    % uhat
shat = uhat>graythresh(uhat);         % estimated spike train

Here is another example with a higher noise level and some bursting:

deconv_hard

Here is how it runs on a some of the cases of the Chen et al dataset that have varying degrees of difficulty.  The red dots represent the spikes detected as described in the original paper, the black dots represents the spikes as detected by the deconvolution algorithm above.

chen_010 chen_009 chen_008 chen_007 chen_006 untitled chen_005 chen_004 chen_003 chen_002 chen_001

Given the simplicity of the algorithm, it performs remarkably well and can be run across many ROIs in real time.

But wait… there is more. One may expect the impulse response to be not purely exponential. In such a case, it may be helpful to generalize the above to a higher-order model. You can then estimate \hat{u}_n in just two lines of Matlab as follows:

a = lpc(y,10); % order=10
uhat = y-filter([0 -a(2:end)],1,y);

The above should bring back memories to anyone who studied speech compression and recognition at some point in time.

The graph below shows what happens when the impulse response is a gamma function (left).  Now, the LPC coefficients are non-zero for lags larger than one.

gamma_lpc

Still, and simple thresholding of \hat{u}_n still recovers the original spike train even under the presence of substantial noise:

gamma_lpc_resultGo ahead, load some of your calcium imaging data and plot both your signal y_n along with \hat{u}_n estimated by the two lines above. Here is are a couple of data segments, again, from the Chen et al dataset.

uhat1 uhat0

Note: There are other pre-processing steps I took (like projecting out the neuropil signal) that I have not described above.

The calculation:

We want to minimize J = E\{( y_n - \alpha y_{n-1} - u_n )^2\}.  Taking the derivative with respect to \alpha and equating to zero leads to

E\{y_n y_{n-1}\} - \alpha E\{y_{n-1}^2\} - E\{u_n y_{n-1}\} = 0

or

m_{12}- \alpha  m_{02} - \lambda \mu = 0

where E\{ u_n\} = \lambda and E\{ y_n \} = \mu.

But taking expectations on both sides of y_n = \alpha y_{n-1} + u_n leads to

\mu = \alpha \mu + \lambda or \lambda = \mu (1-\alpha).

Substituting \lambda in the equation above and solving for \alpha gives \alpha = (\mu^2 - m_{12})/(\mu^2 - m_{02}).