Neurite tracking in time-lapse calcium images using MRF-modeled pictorial structures

Locomotion is controlled by sensory neurons, yet a fundamental, unanswered question is how this happens. We use time-lapse calcium imagery to observe and model the link between locomotion and sensory neuron activity in larva Drosophila. The main issue with calcium images is that they produce low responses when there is little or no neuronal activity. In this work, we use a neurite centerline tracking approach that tackles this issue, even under significant deformations during movement. It incorporates image appearance, local and global shape, as well as motion, in a Markov Random Field (MRF) framework. The objective function is optimized using quadratic pseudo-boolean optimization with a-expansion, which is also known as fusion moves. In our experiments we illustrate how our method can track neurites under severe local intensity ambiguities.