SmartTracing: self-learning-based Neuron reconstruction
Brain Informatics
SmartTracing: self-learning-based Neuron reconstruction
Hanbo Chen 0 1 2
Hang Xiao 0 1 2
Tianming Liu 0 1 2
Hanchuan Peng 0 1 2
0 H. Chen T. Liu Cortical Architecture Imaging and Discovery Lab, Department of Computer Science and Bioimaging Research Center, The University of Georgia , Athens, GA , USA
1 H. Chen (&) H. Peng Allen Institute for Brain Science , Seattle, WA , USA
2 H. Xiao CAS-MPG Partner Institute for Computational Biology, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences , 320 Yueyang Road, Shanghai , China
In this work, we propose SmartTracing, an automatic tracing framework that does not require substantial human intervention. There are two major novelties in SmartTracing. First, given an input image, SmartTracing invokes a user-provided existing neuron tracing method to produce an initial neuron reconstruction, from which the likelihood of every neuron reconstruction unit is estimated. This likelihood serves as a confidence score to identify reliable regions in a neuron reconstruction. With this score, SmartTracing automatically identifies reliable portions of a neuron reconstruction generated by some existing neuron tracing algorithms, without human intervention. These reliable regions are used as training exemplars. Second, from the training exemplars the most characteristic wavelet features are automatically selected and used in a machine learning framework to predict all image areas that most probably contain neuron signal. Since the training samples and their most characterizing features are selected from each individual image, the whole process is automatically adaptive to different images. Notably, SmartTracing can improve the performance of an existing automatic tracing method. In our experiment, with SmartTracing we have successfully reconstructed complete neuron morphology of 120 Drosophila neurons. In the future, the performance of SmartTracing will be tested in the BigNeuron project (bigneuron.org). It may lead to more advanced tracing algorithms and increase the throughput of neuron morphology-related studies.
SmartTracing; Neuron reconstruction; Neuron morphology; Machine learning; Reconstruction confidence
1 Introduction
The manual reconstruction of a neuron’s morphology has
been in practice for one century now since the time of
Ramo´n y Cajal. Today, the technique has evolved such that
researchers can quantitatively trace neuron morphologies
in 3D with the help of computers. As a quantitative
description of neuron morphology, the digital
representation has been widely applied in the tasks of modern
neuroscience studies [
1–3
] such as characterizing and
classifying neuron phenotype or modeling and simulating
electrophysiology behavior of neurons. However, many
popular neuron reconstruction tools such as Neurolucida
(http://www.mbfbioscience.com/neurolucida) still rely on
manual tracing to reconstruct neuron morphology, which
limits the throughput of analyzing neuron morphology.
In the past decade, many efforts have been given to
eliminate such a bottleneck by developing automatic or
semi-automatic neuron reconstruction algorithms [
1, 3
]. In
these algorithms, different strategies and models were
applied, such as pruning of over-complete neuron trees
[
4, 5
], shortest path graph [
6
], distance transforms [
7
],
snake curve [
8
], and deformable curve [
9
]. However, the
completeness and the attribute of resulted neuron
morphology vary tremendously between different algorithms.
Recently, to quantitatively assess such variability between
algorithms and advance the state of the art of automatic
neuron reconstruction method, a project named BigNeuron
[
10, 11
] has been launched to bench-test existing
algorithms on big dataset. One reason causing such variability
is that image quality and attributes vary between different
data sets—partially due to the differences in imaging
modality, imaging parameter, animal model, neuron type,
tissue processing protocol, and the proficiency of
microscopic operator. And some of the algorithms were
developed based on specific data or were developed to solve
specific problem in the data which may not be applicable
for other types of data. Another reason is that most of the
tracing algorithms required user input of parameters. As a
consequence, the optimal parameters vary between images
and thus require manual tuning by the user with sufficient
knowledge of the algorithm.
We note that most of the current automatic neuron
reconstruction algorithms are not ‘‘smart’’ enough. Indeed,
many times they require human intervention to obtain
reasonable result. To conquer this limitation, one can adapt
learning-based methods; so the algorithm can be trained for
different data. In [
12
], the authors proposed a machine
learning approach to estimate the optimal solution of
linking neuron fragments. However, the fragments to link
were still generated by model-driven approaches, and it
requi (...truncated)