Neuron recognition with hidden Neural Network Random Fields

We model neuron morphology with a hidden Conditional Random Field variant, a hidden Neural Network Random Field, for part-based classification. We aim at identifying the diverse morphologies of individual motor neurons in the Drosophila larvae, and understanding underlying principles of synaptic connectivity in a motor circuit. The motivation of our work is the bottom-up reconstruction of the Drosophila connectome, ie., fully annotated neuronal circuits where neurons and synapses are automatically not only traced but also identified. We use images depicting single neurons labeled with green fluorescent protein (GFP) and serially imaged with laser scanning confocal microscopy. In our approach, we consider that each neuron has already been partitioned into its structurally significant parts, namely soma, axon, and dendrites, and their (latent) labels are known. We demonstrate the accuracy of our approach using wild-type motor neurons in the larval ventral nerve cord, and make comparisons with existing methods, including our previous work.