Semi-Automated Neuron Tracing with Hidden Markov Modeling

Like Comment
Read the paper

To fully understand how the brain works, and deviates under disease, the global scientific community is generating comprehensive brain circuit diagrams, which detail how cells in the brain are connected to one another. Currently, this requires human researchers to examine brain images and trace neuron trajectories. This process is extremely tedious, as neurons can branch dozens of times, are only about a micron thick but can extend to hundreds of millimeters in length. In the past decade, several automated neuron tracing methods have been proposed, but they are typically limited to small images, or do not generalize well to different image modalities.

The neuron tracing problem is essentially an image segmentation problem, where there is an object of interest (neuron(s)) that needs to be extracted from a larger image (brain image). Researchers in machine learning have produced impressive image segmentation models that have contributed to self-driving technology and software that can automatically tag faces in social media posts. The long and thin morphology of neurons, however, means that a small segmentation error can completely sever neurons, or fuse them.

Athey et. al. designed a method that starts with an imperfect image segmentation, then proceeds to connect neuron fragments that compose a neuron branch. Their approach is based on hidden Markov modeling (HMMs), which became hugely popular in the 1980s in field of speech recognition. In HMMs, the observed variables (e.g. audio signal) are used to infer the hidden variables (e.g. words being spoken). These methods typically leverage patterns of the hidden variables (e.g. “the” often begins sentences, but rarely ends them). In the neuron tracing context, the neuron trajectory (hidden variable) is being estimated according to the image data (observed data) and typical neuron geometry. The algorithm is called ViterBrain as a reference to the famous Viterbi algorithm in HMMs.

The method has been shown to outperform state-of-the-art neuron tracing algorithms on a dataset of partial axons from the MouseLight project at HHMI Janelia. The implementation is still under development to improve its accuracy, and scalability. It has been incorporated into a prototype graphical user interface, and is freely available in our open-source Python package brainlit.

Thomas Athey

Graduate Student, Johns Hopkins University