Figure 1: Illustrated steps of maskNMF pipeline.

(A) We start with a motion-corrected dataset. (B) We then compress and denoise the data using penalized matrix decomposition (PMD) from Buchanan et al. (2018). (C) Then, we perform pixel-wise deconvolution to temporally sparsen the data; here, we show the result of temporally sparsening a single pixel of the denoised data. (D) We then project this sparse video back onto the spatial denoising basis $U$; note that the resulting image contains a single, well-isolated neural shape, unlike (B). (E) Finally, we run a specialized neural network (a Mask R-CNN architecture, trained on simulated calcium imaging data) to detect neuron signals present in this frame of the data.