Review of advances and prospects in neuron reconstruction

Abstract

Neuronal reconstruction, which involves generating the quantitative datasets that can measure neuronal morphologies, is a key step towards identifying cell types, building neuronal connectivity pattern, mapping neuronal circuits, modeling neurons, and realizing other such activities. Neuronal reconstruction can also provide valuable datasets for some studies focused on artificial intelligence. Recent advances in light microscopic techniques have enabled the whole-brain imaging of a mouse at cellular resolution and allowed us to observe nearly the complete morphology of a population of neurons including dendrites, soma, and axons. These imaging techniques have promoted the study of neuronal networks to an unprecedented scale and have been initiated to obtain the organization of the entire brain at a single-cell level. However, the neuronal datasets collected with these imaging techniques are complicated and pose many challenges to the current neuronal reconstruction methods. In this circumstance, we elaborated upon the research status of neuronal reconstruction, reviewed typical methods, and predicted the research trends. The research status indicates that many current methods focus on single neuron reconstruction, which involves analyzing the images in which a single neuron is included. These single neuron reconstruction methods are classified into groups including skeleton methods, region growing methods, graph-based methods, snake models, and local structure-based methods. We summarized the advantages and disadvantages of each of these methods, and regarded that the reconstruction performance depends on the match between the data set characteristics and the reconstruction methods. A reconstruction method may be suitable for most datasets but fail for some specific cases. In comparison to single neuron reconstruction, neuronal population reconstruction should be paid more attention to: Unlike single neuron reconstruction, neuronal population reconstruction focuses on images including a population of neurons and requires assigning the traced neurites into their own neurons. Neuronal population reconstruction can provide the information necessary for researching the correlation, coupling, and connectivity between neurons, and has several advantages over single neuron reconstruction. We presented compelling evidence to suggest that it is inevitable for the trend of development to move from single-neuron to neuronal population reconstruction. The evidences can be summarized as follows: The existing imaging techniques produce a large amount of neuronal population datasets. However, at present, single neuron reconstruction methods cannot cope with these datasets; population reconstruction method is in its infant stage and is thus urgently required. In addition, many existing methods for single neuron reconstruction have become fully mature and can vastly inspire population reconstruction methods. Furthermore, we highlighted the bottleneck problem of population reconstruction at the brain-wide scale. The current automatic methods fail to cope with this case while semi-automatic reconstruction is too slow. Using the conventional methods, such as a combination of big data organization techniques and the Amira tracing module, thousands of hours are required for the reconstructions of several tens of brain-wide neurons. We proposed a strategy for this bottleneck problem as designing automatic algorithms for the laborious steps in this reconstruction. Finally, we expressed our views on the future development of the neuronal population reconstruction. This sort of reconstruction has a complicated pipeline that involves many steps. Aiming at some key steps, machine learning methods will be designed. The above contents will help researchers evaluate the latest developments in the field at a glance and develop new reconstruction methods that match with the characteristics of the current datasets.