With the development of optical imaging systems, neuroscientists can now obtain large datasets of morphological structure at a single neuron scale positioned across the whole mouse brain. However, the enormous amount of morphological data challenges the classic approach of neuron classification, indexing and other analysis tasks. In this paper, we propose MorphoGNN, a single neuron morphological embedding based on the graph neural networks (GNN). This method learns the spatial structure information between the nodes of reconstructed neuron fibers by its nearest neighbors on each layer and captures the lower-dimensional representation of a single neuron through an end-to-end model. This model is composed of densely connected edge convolution (EdgeConv) layers and a double pooling operator, regularized with joint cross-entropy loss and triplet loss. An increasing population of the neighbor nodes meets the need of learning more information with features expanding at the deep layer. We tested the proposed embeddings on the neuron classification and retrieval tasks. Our method achieves competitive performance both on the general point cloud dataset and the neuron morphology dataset.

No abstract

No abstract