Neurobfd: Size-Independent Automated Classification Of Neurons Using Conditional Distributions Of Morphological Features

Automated classification and characterization of digitally-traced neurons poses a challenge in constructing the neurome of human brain. The complex anatomical structure of each neuron causes observable morphological and geometrical variations between different cell types and within cell types. Given a graph model of a neuron, one of the major bottlenecks in encoding such variation followed by the within-cell and between-cell comparisons is the size of the neuron. In this work, we define size-independent statistical morphometrics as feature descriptors for each neuron. The customized morphometrics are built by extracting three raw features, which are bifurcation angle, branch fragmentation, and spatial arborization density of a neuron. Next, the local variation in each of the raw features is encoded by constructing a conditional distribution for that feature, which provides an effective and discriminatory feature assembly. In addition, the algorithm is automatic, scalable, and coordinate-independent. The comparative approach is shown to outperform the existing state-of-the-art methods.