Manifold classification of neuron types from microscopic images

Abstract

Analysis of cell types is recognized as a major task in current single-cell genotyping and phenotyping. In neuroscience, 3-D neuron morphologies are often reconstructed from multi-dimensional microscopic images. Recent studies indicate that neurons could form very complicated distributions in the feature space, and thus they can be explored using manifold analysis. We have developed manifold classification toolkit software to replace the conventional clustering analysis to discover cell subtypes from three state-of-the-art collections of single neurons’ 3-D morphologies that reconstructed from images. We have gathered 9208 3-D spatially registered whole mouse brain neurons from three datasets with the highest quality to date generated by the single neuron morphology community. To explore manifold distribution, our method uses minimum spanning tree-based principal skeletons to approximate locally linear embeddings, to explore the morphological feature spaces, which correspond to dendritic arbors, axonal arbors or both categories of arborization patterns of neurons. We show manifold classification is a suitable approach for a majority of often referred cell types, each of which was also discovered to contain multiple subtypes. Our results show an initial effort to employ manifold classification but not traditional clustering analysis as an alternative framework for analyzing 3-D neuron morphologies reconstructed from 3-D microscopic images.