Fluorescent calcium indicators are a widely employed technique to study neural activations. However, the chemical nature of the reaction can be problematic, as fluorescent traces have a slow decay, compared to the underlying electrical activations. In this work, we present a Bayesian deconvolution model to extract the time series of neural activations, the spike train, from the signal. To ensure the required sparsity, we employ a generalised horseshoe prior distribution within the framework of a structural time-series model. We show that the approach can be coded into any probabilistic programming languages and give an example with Stan.