Automatic 3D Neuron Tracing from Optical Microscopy Images.

Neuron tracing is the process of reconstructing three-dimensional morphology of neurons from microscopy images. It is essential for delivering more comprehensive understanding of the relationship between neuronal structure and function, which is the fundamental to know how the brain works. However, currently neuron tracing remains a challenging task, due to the natural complexity of neuronal structure, inadequate available data and computational limitation. In recent years, many automatic neuron tracing methods have been developed in the research field, with limited success on specific issues. The lack of a robust neuron tracing method with more general applicability greatly restrains systematic characterisation and analysis on neuronal morphology. To address aforementioned challenges, we first establish a pipeline to generate more standard data, in which we specifically propose a novel approach for automatic refinement on semi-manual reconstruction. Following the pipeline, we manage to generate more than 1000 full morphology data. Second, based on the generated standard reconstruction, we conduct a systematic and quantitative analysis to identify the most critical obstacles in neuron tracing. Third, we propose a novel neuron tracing method by embedding occupancy learning with curve skeleton extraction, which tackles the major issues of weak and punctuated signal, as concluded from the previous analysis. We curated a large dataset to train and test the model. The experimental results show it exceeds other counterpart approaches in most terms of evaluation metrics. At last, we propose a novel learning model for automatic neuron tracing, which learns to directly extracts the skeleton from a raw image. It addresses the main issue of close but irrelevant signal, as concluded previously. We train and bench test it on the curated dataset, as well as a public dataset. Experiments show it achieves state-of-the-art performances in all cases.

Level dc:type.qualificationlevel

doctoral

Grantor dc:publisher.institution

Bournemouth University

Year dc:date.issued

2022

Author dc:creator

• Ding, Zhangcan

Language dc:language

en