INFORMATION TECHNOLOGY

Break through the challenge and achieve automated and accurate measurement of neurons


The human brain is a complex network of tens of billions of neurons connected. The development, benchmarking and accuracy measurement of large-scale neuronal morphological reconstruction algorithms are of great significance for understanding the development and function of the brain and the diagnosis and treatment of neurological diseases.

Recently, the Institute of Brain Science and Intelligent Technology of Southeast University has joined hands with a number of domestic and foreign teams to make important research progress in the benchmarking and performance prediction of automatic neuron tracking algorithms, and the relevant results have been published in the top international methodology journal “Nature – Methods”.

A huge challenge

The brain is a very important organ, containing 86 billion neurons, and it is these neurons that function like a sea of stars.

“To gain insight into the function and principles of the brain, we need to first understand the morphology of neurons. However, obtaining the complete morphology of neurons is very difficult because it requires a lot of time and labor to manually annotate. One possible approach is to use automated algorithms for neuronal morphological reconstruction, but there is currently a lack of standard datasets and accuracy evaluation methods. Liu Yufeng, a co-author of the paper and a doctor of Southeast University, said in an interview with China Science News.

The structure of neurons is very tiny, usually on a submicron scale, while mammalian neurons often span hundreds of millimeters or even centimeters. Therefore, to rebuild a complete neuron, it is necessary to have large-scale and high-resolution data, which leads to a very large amount of data. Even for mouse brains, about 50 terabytes of data need to be processed, which is a huge challenge for the algorithm.

So far, only a few thousand intact mammalian (mouse) neurons have been reconstructed. In a 2019 article published in Cell, Janellia Farm’s team reconstructed 1,000 neurons. Then, in 2021 and 2022, the research team led by the Southeast University-Allen Research Center and the Center for Excellence in Brain Science and Intelligent Technology of the Chinese Academy of Sciences successfully reconstructed 1741 complete neuronal morphology and 6357 projected neurons, respectively, and the results were published in the journals Nature and Nature Neuroscience.

The study of neuronal diversity, the connection between morphology and function, and the mapping of brain connections at the neuron level all require high-throughput neuronal tracking techniques. This is especially important in large-scale data brain projects, such as the China Brain Program and the American Brain Program.

“However, neuronal tracking in light microscopy datasets is still manual and labor-intensive. Although many automated tracking algorithms have been proposed, the results used are frustrating because the quality of results for different methods varies greatly under different imaging conditions. Linus Manubens-Gil (Chinese name Lin Jier), co-first author of the paper and a researcher at the Institute of Brain Science and Intelligent Technology of Southeast University, said in an interview with China Science News.

One challenge in the field is the development of fully automated neuronal tracking tools.

But over the past 40 years, attempts to develop such algorithms have shown that this is a huge challenge, because existing optical microscopic images often contain a lot of noise and blemishes, which can easily lead to automatic neuronal reconstruction errors. And even single-node errors, such as missing a branch point in the axon of a neuron, can lead to unsatisfactory reconstruction. So this is a very sensitive issue that requires highly accurate algorithms.

“While many research teams have come up with auto-tracking algorithms, tools that perform well on a particular dataset, such as those generated in one lab, are not applicable to other teams’ research areas.” So another challenge, Linus Manubens-Gil says, is benchmarking existing tools that define open standards.

An overview of BigNeuron and how it interacts.  Photo courtesy of interviewee

A multifaceted study

The research team collected and shared a dataset of approximately 30,000 cross-species, widely diverse three-dimensional neuronal images, and used them to benchmark the automated tracking algorithm, generating more than 1.4 million neuronal reconstructions, resulting in the largest neuronal reconstruction benchmark dataset to date.

“For a research field, standardization is the most basic and difficult step, once the data and methods are standardized, the latter things are relatively much easier, and ‘talking to yourself’ is avoided.” The researchers say that there is currently a lack of high-quality test data (“gold standard”) and test methods in the industry, and their study solves this problem.

The research team selected 166 neuronal images from neuronal reconstruction and manually annotated their “gold standard” reconstruction patterns for posterior benchmarking of the automatic tracking algorithm.

As part of this research, the research team also developed an interactive web application that enables users and developers to visualize image data and neuron reconstruction results and perform various analyses on them. An interesting phenomenon is that different algorithms can provide complementary information, so they developed a method of iteratively combining different algorithms to produce consistent reconstruction results, improving the accessibility, accuracy, and efficiency of automatic reconstruction methods.

Another contribution to the study is the development of a new method for predicting the accuracy of automated neuron reconstruction algorithms, using a specific metric of image quality and a set of reconstruction results as input, without the need to manually annotate images.

“Currently, BigNeuron-annotated datasets and benchmark algorithms have been used in hundreds of studies, and there is no doubt that it will continue to be a major resource for researchers developing new algorithms or determining which existing algorithms are best suited for their experiments.” Erik Meijering, co-corresponding author of the paper and a professor at the University of New South Wales in Australia, said.

All in all, this study sets a standard for addressing various technological hurdles, with the ultimate goal of understanding how individual neurons and neuronal networks function in health and disease.

A friendly competition

The reason why this research was carried out smoothly was due to the cooperation of researchers from different disciplines such as computer science, neuroscience, and neuroinformatics from many countries. The number of authors of this paper is as high as 65.

Moreover, the collaborative approach is particularly unique, a critical shift from competitive to collaborative to solve open problems in neuroscience.

Over the past decade, research teams have organized many international competitions in the field of computer vision and biomedical image analysis with the aim of objectively comparing the performance of state-of-the-art algorithms in solving various problems.

But beyond answering the question ‘who is currently the best?’ most of these studies have done little to the field in terms of creating opportunities and fostering collaboration. Some even refer to such studies as “fight clubs.”

In contrast, the BigNeuron project has been a “team fight” from the start, “encouraging computer scientists, neuroscientists, and neuroinformaticians to work together in various hackathons and workshops organized over the years.” This is evidenced by the long list of authors in the paper. Erik Meijering said.

The research team benchmarked the single-neuron autotracking algorithm on the universal open platform Vaa3D, held a series of hackathons and events, developed 16 auto-tracking algorithms, and quantified the reconstruction quality of 35 auto-tracking algorithms (16 algorithm variants) with “gold standard” data as a reference.

It’s a friendly competition. This is in stark contrast to the traditional reward-based mechanism of supporting neuroscience research, which inevitably pits labs against each other.

“I hope that our research will inspire the development of new algorithms, and that our tools will facilitate neuronal tracking in light microscopy.” Linus Manubens-Gil said. (Source: Zhang Qingdan, China Science News)

Related paper information:https://doi.org/10.1038/s41592-023-01848-5



Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button