Automated Rat Neuronal Morphology Analysis via Self-Supervised Graph Networks

Neuronal morphology is a key indicator for describing neuronal cell types, analyzing brain developmental processes, and diagnosing changes in neurological diseases. The advancement of neuronal labeling, imaging and reconstruction techniques has greatly expanded neuronal data and promoted neuronal morphology. However, effectively analyzing the extensive and diverse 3D neuronal data remains a considerable challenge. The traditional analysis of neurons mainly relies on manual defined features and visual assessments. Due to the complexity and diversity of neuron data, the above methods are not sufficient to distinguish the fine-grained features in a large number of neurons. In this paper, we address the above problems by introducing neuron morphological characterization algorithm based on graph neural networks through self-supervised learning, which is capable of obtaining a superior neuron characterization. Comparison with the prior methods shows that the algorithm in this paper outperforms the existing supervised algorithms even without annotations.