A number of sophisticated deconvolution algorithms have been proposed in the literature, most of which involve solving an 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, , is the response of a linear system with an exponential impulse response to a train of impulses .
Then, the response satisfies , where we assume is Gaussian noise.
One simple way to estimate is to first find the value of that minimizes the error .
A few calculations yield where , and .
As all these quantities are easily estimated recursively as new data arrive, one can also maintain a recursive estimate of the optimal .
Then, the estimated input to the system can be estimated by computing . So, all we need to do is to keep a recursive estimate of and the last two samples of the signal.
Below is an example of how this simple algorithm performs.A simulated spike train, (second panel form top) is convolved with an exponential impulse response and corrupted by additive Gaussian white noise to yield a simulated signal (top panel). The third panel shows the estimated . 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)&lt;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&gt;graythresh(uhat); % estimated spike train
Here is another example with a higher noise level and some bursting:
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.
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 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.
Still, and simple thresholding of still recovers the original spike train even under the presence of substantial noise:
Note: There are other pre-processing steps I took (like projecting out the neuropil signal) that I have not described above.
We want to minimize . Taking the derivative with respect to and equating to zero leads to
where and .
But taking expectations on both sides of leads to
Substituting in the equation above and solving for gives .