Comparing neurons using a direct method from computer vision
In the lab, we are generally interested in how vision works within the brain and which roles the individual types of neurons have. While previous studies on relatively few cells (hundreds) have indicated a number of cell types, we now have a much bigger dataset of high-quality neuron scans available (tens of thousands). This raises the question in how far data-driven models reproduce the expert-generated neuron classes. In the bigger scheme of things, being able to compare and cluster neurons based on their shape is a step towards answering the more general question in how far the shape of a neuron tells us something about the function of that neuron.
In this project, the task is to use a technique from computer vision, originally created to morph one object into another in a visually pleasing way, to compare 3D scans of mouse brain neurons. The method itself uses a “longest-path-decomposition” of the neuron and represents each path by a function as this allows it to stretch and bend the path to match another neuron. It then creates a matching of its subtrees and moves them to their new location. The distance between two neurons then consists of a weighted sum of the “bending and stretching” as well as the “moving of subtrees”. We are interested how this method compares to the fully data-driven metrics we are currently using to compare and cluster neurons.
Possible extensions:
- Creating convincing averages of neuron clusters identified by other means (the basis is given in the paper)
- Optimizing the code to make it scale to larger datasets (from one hundred to thousands)
- Porting the method to use the GPU (and python), in case the computed distances are meaningful for the neuroscience task
Prerequisites:
- Programming experience
- Matlab experience is a bonus (the code is written in matlab)
Literature
Paper: https://arxiv.org/abs/2110.08693v2
Github: https://github.com/Fanta007/complexTrees_code
