maskNMF: a denoise-sparsen-detect pipeline for demixing dense imaging data faster than real time

In recent years, many calcium and voltage imaging methods have been developed in pursuit of a central goal of systems neuroscience: to study the function of large populations of neurons. Projective imaging methods, such as Bessel imaging, are a promising class of approaches in pursuit of this goal. These methods simultaneously record neural activity at a wide range of depths in the brain, effectively allowing scientists to image a volume of brain tissue. However, this added capability comes with a trade-off: all observed signals from the volume are projected onto a 2D imaging plane. Therefore, these signals must be computationally demixed to recover the desired neural activity. Furthermore, these large-scale recordings demand computationally fast analyses.

We address these needs with a GPU accelerated algorithm, maskNMF, which can demix hour-long Bessel data within minutes. The key strategy in maskNMF is to first estimate the spatial profiles of the neurons, and use these estimates as a starting point for spatially and temporally demixing the entire dataset. To that end, we begin by registering and compressing the imaging video. This has the added benefit of denoising the data. Next, we use a deconvolution operation on the compressed data to temporally sparsen the imaging video. The result typically contains spatially isolated neurons in each frame. We use a neural network trained automatically on simulated calcium imaging data to detect these isolated neural shapes.

We then use these spatial shapes to initialize a constrained nonnegative matrix factorization model (NMF) to perform signal demixing. Our method can accurately demix denser data than was previously feasible, enabling imaging of larger neural populations. We have validated this method on both real and simulated data and provide a detailed comparison of how state-of-the-art methods compare to maskNMF as imaging density increases.