3D Voxel Model for Representation Learning of Neuronal Morphologies
Background
The morphology of neurons is highly complex with widely varying shapes that are tightly linked to their function. Traditionally, neurons have been classified based on their morphology into diverse types using manual inspection. Recent advances in recording technologies have greatly accelerated data collection and therefore the amount of data available. These developments provide optimal conditions for applying unsupervised machine learning methods to characterize neuronal morphologies and identify cell types.
Master’s thesis
The aim is to apply a voxel-based model [1,2,3] to embed the 3D morphologies of the neurons into low-dimensional vector embeddings — also referred to as “bar codes” — for each neuron. The latent space of those bar codes should be organized in a way that we can potentially sample and generate new, biologically realistic neurons. Further, clustering those bar codes gives us information about similar and distinct neurons and can be compared to functional clusterings in a second stage. We work on optimizing such models to gain a better understanding of the neurons' morphologies and, in the future, potentially link it to their function.
Requirements
- Python programming
- Interest in 3D data
- Interest in working with neurons and biological data
- Experience with PyTorch
Literature
[1,2] (Scalable) SoftGroup for 3D Instance Segmentation on Point Clouds
[3] PVT: Point-Voxel Transformer for Point Cloud Learning
