Calcium imaging has emerged as a workhorse method in neuroscience to investigate patterns of neuronal activity in vivo. Still, algorithms to automatically detect and extract activity signals from calcium imaging movies are highly variable from lab to lab and more advanced algorithms are continuously being developed.
We have developed HNCcorr, a novel algorithm for cell identification in calcium imaging movies based on combinatorial optimization. HNCcorr guarantees an optimal solution and has minimal dependence on initialization techniques.
The algorithm identifies cells by finding distinct clusters of highly similar pixels. HNCcorr uses a new method for computing similarities named $\text{(sim})^2$, similarity squared. The idea of $\text{(sim})^2$ is to associate with each pixel a vector of correlation similarities with respect to a reference set of pixels, and determine the similarities between pairs of pixels by computing the similarity of the respective two vectors. HNCcorr achieves the best known results for the Neurofinder cell identification benchmark.