Automating the Reconstruction of Neuron Morphological Models: the Rivulet Algorithm Suite

Abstract

The automatic reconstruction of single neuron cells is essential to enable large-scale data-driven investigations in computational neuroscience. The problem remains an open challenge due to various imaging artefacts that are caused by the fundamental limits of light microscopic imaging. Few previous methods were able to generate satisfactory neuron reconstruction models automatically without human intervention. The manual tracing of neuron models is labour heavy and time-consuming, making the collection of large-scale neuron morphology database one of the major bottlenecks in morphological neuroscience. This thesis presents a suite of algorithms that are developed to target the challenge of automatically reconstructing neuron morphological models with minimum human intervention. We first propose the Rivulet algorithm that iteratively backtracks the neuron fibres from the termini points back to the soma centre. By refining many details of the Rivulet algorithm, we later propose the Rivulet2 algorithm which not only eliminates a few hyper-parameters but also improves the robustness against noisy images. A soma surface reconstruction method was also proposed to make the neuron models biologically plausible around the soma body. The tracing algorithms, including Rivulet and Rivulet2, normally need one or more hyper-parameters for segmenting the neuron body out of the noisy background. To make this pipeline fully automatic, we propose to use 2.5D neural network to train a model to enhance the curvilinear structures of the neuron fibres. The trained neural networks can quickly highlight the fibres of interests and suppress the noise points in the background for the neuron tracing algorithms. We evaluated the proposed methods in the data released by both the DIADEM and the BigNeuron challenge. The experimental results show that our proposed tracing algorithms achieve the state-of-the-art results.