Estimating the posture of coiled and overlapping worms using a neural network

The first step in studying postural dynamics is typically animal tracking. A range of deep learning approaches have proven effective for human posture tracking and several groups have adapted these methods to laboratory animals. These methods typically detect key points on the body and they work particularly well for jointed animals where there are clear landmarks. For tracking the nematode worm C. elegans, some challenges remain. We generated a manually annotated data set with tens of thousands of frames including coiled and overlapping worms to train a neural network to find equidistant points along the worm midlines. During training we rely on simple image synthesis augmentations and use priors that incorporate the worm morphology into the loss function. We show that the new algorithm performs well with challenging backgrounds, for coiled worms, and for multiple overlapping worms. We apply the network to mating, a behaviour that necessarily involves overlapping worms.