Variational Autoencoder with Temporal Condition for Effective Shape-based Calcium Imaging Neuron Registration.

Thanks to recent advances in optical imaging techniques, calcium imaging can now record the activities of thousands of neurons simultaneously, through several sessions and over long periods. Neuron registration assumes a vital role in this process, enabling the monitoring of neurons across multiple movies and extended timeframes by aligning their spatial patterns. Previous approaches often relied on clusters or probabilistic models based on simple distance metrics like overlapping pixels or center-to-center distances, neglecting crucial neuron shape and spatial relationship details. In this paper, we introduce a novel technique for cell registration. Our investigation revealed that a neuron’s shape is influenced by its temporal behaviors, leading us to suggest a temporal-conditional variational autoencoder (tcVAE) for precise shape modeling. A comprehensive evaluation demonstrates that incorporating shape-related details can significantly enhance the quality of neuron registration.