Artificial intelligence selects neurons faster than the human

Biomedical engineers at Duke University have developed an automated process that can trace the shape of active neurons as accurately as human researchers can, but in a fraction of the time.

This new technique, based on the use of artificial intelligence to interpret video images, addresses a major obstacle in neural analysis, allowing researchers to quickly gather and process neural signals for real-time behavioral studies.

The research appeared this week in the Proceedings of the National Academy of Sciences.

To measure neuronal activity, researchers typically use a process called two-photon calcium imaging, which allows them to record the activity of individual neurons in the brain of living animals. These recordings allow researchers to know which neurons are activated and how they potentially correspond to different behaviors.

Although these measurements are useful for behavioral studies, the identification of individual neurons in the recordings is a tedious process. Currently, the most accurate method requires that a human analyst surrounds each "spark" that he detects in the recording, often asking them to stop and rewind the video until they are heard. Targeted neurons are identified and saved. To further complicate the process, researchers are often interested in identifying a small subset of active neurons that overlap in different layers within thousands of imaged neurons.

This process, called segmentation, is complex and slow. A researcher can spend four to 24 hours segmenting the neurons into a 30-minute video recording, assuming they are perfectly focused for the duration and that they do not take a break to sleep, eat or use the bathroom.

In contrast, a new open source automated algorithm developed by researchers in image processing and neuroscience at Duke's Biomedical Engineering Department can accurately identify and segment neurons in minutes.

"As a crucial step in the complete mapping of brain activity, we have been tasked with developing a fast automated algorithm, as accurate as humans, to segment a variety of neurons. assets imaged in different experimental contexts, "said Sina Farsiu. Paul Ruffin Scarborough Associate Professor of Engineering at Duke BME.

"The bottleneck in data analysis has been around for a long time in neuroscience, and analysts have spent hours and hours processing data minutes, but this algorithm can handle a 30-minute video. 20 to 30 minutes, "said Yiyang Gong, assistant professor. in the Duke BME. "We have also been able to generalize its performance so that it can work as well if it is necessary to segment the neurons of another layer of the brain with a different size or density of neurons."

"Our algorithm based on deep learning is fast and it is shown that it is as accurate as (if not better than) human experts in the segmentation of active and overlapping neurons from within the world." two-photon microscope recordings, "said Somayyeh Soltanian-Zadeh, Ph.D. student at Duke BME and first author on paper.

In-depth learning algorithms allow researchers to quickly process large amounts of data by sending them across multiple layers of non-linear processing units, which can be trained to identify different parts of a data. complex picture. In their framework, this team has created an algorithm capable of processing both spatial and temporal information in input videos. They then "trained" the algorithm to mimic the segmentation of a human analyst while improving accuracy.

This breakthrough is a crucial step in enabling neuroscientists to monitor neuron activity in real time. Due to the widespread utility of their tool, the team has made its annotated software and dataset available online.

Gong is already using the new method to study more closely the neuronal activity associated with different behaviors in mice. By better understanding the neurons that trigger different activities, Mr. Gong hopes to learn how researchers can manipulate brain activity to alter behavior.

"This improved performance in active neuron sensing should provide more information about the neural network and behavioral states, and pave the way for accelerated progress in neuroscience experiments," Soltanian-Zadeh said.