A cutting-edge AI method developed by researchers at EPFL and Harvard utilizes a convolutional neural network with “targeted enhancement” to effectively monitor cells in moving organisms. This innovation accelerates brain imaging research, minimizes manual annotation, and enhances our understanding of neurological behaviors.
EPFL and Harvard scientists pioneer an AI-driven approach to track neurons in mobile animals, revolutionizing brain research with minimal manual labeling.
Recent advancements have enabled the observation of cells in freely moving organisms. However, accurately identifying and tracking these neurons for deciphering circuit activity poses a significant challenge. This task becomes especially daunting when studying the dynamic and deformable nature of the brain in mobile creatures like worms, presenting an ongoing obstacle for the scientific community.
Introducing an AI solution for neural mapping
A collaborative team of researchers from Harvard and EPFL has introduced an innovative AI technique for recording neurons in animals undergoing movement or deformation. Led by Sahand Jamal Rahi, the study, recently published in Nature Techniques, was conducted at EPFL’s School of Basic Sciences.
At the core of this novel approach lies a convolutional neural network (CNN), a type of artificial intelligence trained to identify and interpret patterns within images. Leveraging a process called “convolution,” the CNN analyzes intricate image details such as edges, colors, and shapes collectively to comprehend and recognize objects or patterns.
The challenge arises from the necessity to manually label certain images of an animal’s brain to detect and record neurons, given the continuous bodily transformations occurring over time. The sheer diversity of postures exhibited by the organism can make it arduous to generate a sufficient amount of annotations to train a CNN effectively.
To address this issue, the researchers devised an advanced CNN incorporating “targeted augmentation.” This state-of-the-art technique autonomously generates reliable annotations from a limited set of initial annotations, significantly reducing the reliance on human intervention and verification. By adeptly learning the internal brain deformations, the CNN can create annotations for new postures with remarkable accuracy.
This adaptable technique enables cell identification whether represented in images as single points or 3D volumes. The researchers tested the method using the model organism Caenorhabditis elegans, a species of tapeworm renowned for its 302 cells.
Employing the enhanced CNN, the researchers monitored the activity within some of the worm’s cells (neurons responsible for transmitting signals between neurons). They observed intricate behaviors, such as modifications in response patterns to diverse stimuli like intermittent bursts of odors.
Implications for research
By introducing a user-friendly graphical interface that integrates targeted augmentation and streamlines the process from human annotation to final validation, the team has democratized the use of CNN technology.
According to Sahand Jamal Rahi, this innovative approach boosts research efficiency threefold compared to traditional full-scale annotation methods by significantly reducing the manual effort required for neuron classification and tracking. This breakthrough has the potential to expedite brain imaging studies and enhance our understanding of neural circuits and behaviors.
