Automated Analysis of Neuronal Morphology: Detection, Modeling and Reconstruction
Dr Daniel Racoceanu, Adjunct Professor, School of Computing
COM2 Level 4
Executive Classroom, COM2-04-02
Abstract:
This thesis addresses the problem of automatic analysis of neuronal tree morphology: from detection, modeling to digital reconstruction of tubular branches of neurites from 3D Light Microscopy image stacks. Comprehending the complex structure and connections of the neurons is key to the study of brain development and functioning. Advances in neuro-imaging has enabled us to capture such neuronal morphology in previously unimaginable details. The huge volume of rich and heterogeneous data generated, however, makes expert manual analysis of such images tedious, subjective and prohibitively expensive. Thus, robust, scalable and highly automated algorithms are desired to analyze such data. In this thesis, we propose a framework for automated quantification of neuronal morphometrics. We target three important methodological areas of biomedical image analysis.
First, we present an automated, unsupervised object detection framework using stochastic marked point processes. We propose special configurations of marked objects and an energy function well adapted for vessel-like tubular structure networks. It incorporates both radiometric properties and high level structural constraints into a probabilistic formulation and global optimization scheme. This enables extraction of connected neuronal networks by fitting an optimal configuration of spheres to the centreline of the branches giving us the position, local width and orientation information.
Second, we explore the problem of modeling of single neuron morphology. We tackle parameter estimation for the marked point process model by relating the model parameters to the application data. A critical analysis of the sensitivity and robustness of the model identifies the parameter dependencies and rules for initialization of the critical parameters. We propose new priors for accurate identification of critical nodes like bifurcations and terminals. Such explicit modeling of neurites derive various characteristic morphological and geometrical parameters such as total length, internodal lengths, branching index, branching angles, average branch curvature etc., to describe the neurons.
Third, we focus on reconstruction of neuronal branches using robust and efficient numerical fast marching methods into connected minimum spanning tree representations. We exploit image potential in evaluating the connectedness of nodes that are Euclidean neighbors to remove the false positives. Thus, we generate a mathematical description abstracting out the important position and connectivity information about neuronal branches from the microscopy data. Such digital reconstruction can be represented in the standard SWC format, prevalent for storage, sharing, and further analysis in the neuroimaging community. Our proposed pipeline outperforms existing neurite tracing algorithms and minimizes the subjective variability in reconstruction, inherent to semi-automatic and semi-manual methods.