Scientists from the Ecole Polytechnique Fédérale de Lausanne (EPFL) and Harvard University have created a groundbreaking artificial intelligence method that uses convolutional neural networks with "directional enhancement" to efficiently track neurons in moving animals. This greatly reduces manual annotation, accelerates brain imaging research, and deepens our understanding of neurobehavior.
Scientists at EPFL and Harvard University have developed an artificial intelligence-based method for tracking neurons in moving animals, improving the efficiency of brain research with minimal manual annotation.
Recent advances allow imaging of neurons in freely moving animals. However, to decode circuit activity, these imaging neurons must be computationally identified and tracked. This becomes particularly challenging when the brain itself moves and deforms within the flexible body of an organism, such as a worm. Until now, the scientific community has lacked the tools to address this problem.
Now, a team of scientists from Ecole Polytechnique Fédérale de Lausanne (EPFL) and Harvard University has developed a pioneering artificial intelligence method for tracking neurons in moving and deforming animals. The research, published in Nature Methods, was led by Sahand Jamal Rahi from EPFL's School of Basic Sciences.
The new method is based on convolutional neural networks (CNN), a type of artificial intelligence trained to recognize and understand patterns in images. This involves a process called "convolution", which looks at small parts of the picture at a time, such as edges, colors or shapes, and then puts all the information together to make sense and identify objects or patterns.
The problem is that to identify and track neurons during the process of photographing an animal's brain, many images must be annotated by hand, because the animals look very different at different times due to different body deformations. Given the diversity of animal poses, manually generating a sufficient number of annotations to train a CNN can be daunting.
To solve this problem, the researchers developed an enhanced CNN with "directional enhancement" function. This innovative technology automatically synthesizes reliable annotations as a reference from only limited manual annotations. As a result, CNNs can effectively learn the brain's internal deformations and then use them to create annotations for new poses, greatly reducing the need for manual annotation and repeated checking.
The new method is versatile and can identify neurons whether they appear as single points in images or as three-dimensional volumes. The researchers tested it on the nematode Caenorhabditis elegans, which has only 302 neurons, making it a popular model organism in neuroscience.
Using enhanced CNNs, the scientists measured the activity of some of the worm's interneurons (neurons that pass signals between neurons). They found that these neurons displayed complex behaviors, such as changing their response patterns when exposed to different stimuli, such as periodic bursts of odors.
The research team made their CNN accessible, providing a user-friendly graphical user interface that integrated targeted enhancements, streamlining the entire process into a comprehensive pipeline from manual annotation to final proofreading.
Sahand Jamal Rahi said: "By drastically reducing the manual effort required to segment and track neurons, the new method increases analysis throughput to three times that of full manual annotation. This breakthrough has the potential to accelerate brain imaging research and deepen our understanding of neural circuits and behavior."
Compiled source: ScitechDaily