Toward reconstructing spike trains from large-scale calcium imaging data

Abstract

Neural activity can be captured by state-of-the-art optical imaging methods although the analysis of the resulting data sets is often manual and not standardized. Therefore, laboratories using large-scale calcium imaging eagerly await software toolboxes that can automate the process of identifying cells and inferring spikes. An algorithm proposed and implemented in a recent paper by Mukamel et al. [Neuron 63, 747-760 (2009)] used independent component analysis and offers significant improvements over conventional methods. The approach should be widely applicable, as tested with data obtained from the mouse cerebellum, neocortex, and spinal cord. The emergence of analysis tools in parallel with the rapid advances in optical imaging is an exciting development that will stimulate new discoveries and further elucidate the functions of neural circuits.

Figure 1. Automated cell sorting and spike train reconstruction from large-scale calcium imaging data.

(A) The workflow of the automated analysis algorithm proposed and implemented by Mukamel et al. (2009). (B) From calcium imaging data recorded in the mouse cerebellum in vivo, the algorithm automatically selected spatial and temporal filter pairs with morphologies and transients expected for individual Purkinje dendrites and Bergmann glia. (C) Simultaneous calcium imaging and single-unit recording confirmed that the extracted fluorescence transients correlated well with spike trains recorded from the two cell types. [Reprinted from Mukamel et al. (2009), with permission from Elsevier].

Figure 2. Automated cell sorting applied to data from the mouse neocortex and spinal cord.

(A) The data set consists of 1000 time-lapse image frames of neocortical cells labeled via multicell bolus loading [H Alitto and Y Dan (2009), unpublished data]. Cell bodies were initially identified manually as ROIs. In most cases, the ICA algorithm identified >90% of the cell in the field of view, although sometimes ICA missed cells with low basal fluorescence or very sparse activity. Moreover, the set of cells selected by the ICA algorithm could vary slightly depending on the number of principal components used. (B) Spinal interneurons were rhythmically bursting during fictive locomotion in mouse spinal cord. The ICA algorithm was able to pick up a subset of the interneurons that have the largest fluorescence changes.