New AI Tracks Neurons in Moving Animals

Source: PixxelTeufel / Pixabay

Scientific research in complex fields such as neuroscience is getting a boost from artificial intelligence (AI) machine learning. A new study by researchers at Ecole Polytechnique Fédérale de Lausanne (EPFL) in Lausanne, Switzerland, and Harvard University in Cambridge, Massachusetts, U.S., shows how AI has the potential to advance neuroscience by identifying and tracking neurons in moving animals.

“Machine learning is ideally suited to automate the task of segmenting and tracking neurons,” wrote lead author Sahand Jamal Rahi, along with coauthors Aravinthan Samuel, Corinne Jones, Vladislav Susoy, Ariane Delrocq, Kseniia Korchagina, Mahsa Barzegar-Keshteli, and Core Francisco Park.

Cognitive neuroscience is the branch of neuroscience and biological psychology that studies the neural mechanisms of cognition. Brain imaging techniques such as functional magnetic resonance imaging (fMRI), electrocorticography (ECoG), magnetoencephalography (MEG), optical imaging with near-infrared spectroscopy (NIRS), and positron emission tomography (PET) are used to study the human brain.

These imaging techniques generate complex, high-dimensional datasets that require segmentation and tracking of the pixels for each neuron. Surprisingly, this often requires painstaking and time-consuming manual annotation. In artificial intelligence, annotation refers to labeling data within datasets to be used by machine learning algorithms. The data elements may be in the form of text, images, videos, or voice data.

To accelerate the annotating of complex brain-imaging data, the EPFL and Harvard researchers used an AI convolutional neural network (CNN) model with the capability to perform targeted augmentation via automating the creation of synthetic annotations from a few manual annotations. The researchers named their technique “Targettrack.”

“We present innovations that allow a CNN to minimize the required amount of manually annotated training data and final proofreading,” the scientists wrote.

Convolutional neural networks are feed-forward deep-learning neural networks commonly used for computer-vision-image classification, speech recognition, audio and signal classification, and natural language processing for text classification.

To conduct their experiment, the scientists used brain-wide imaging data of C. elegans (Caenorhabditis elegans). This nematode worm is often used for research for neurodegenerative diseases, such as Alzheimer’s disease, Parkinson’s disease, Huntington’s disease, and others, as well as ischemia, stroke, age-related diseases, mitochondrial diseases, and immune system responses.

With Targettrack, the AI researchers reported reducing a 200-hour manual task of annotating 76 neurons in a 5 Hertz 10-minute recording of actively moving C. elegans to just 65 hours—a 67.5 percent improvement. According to scientists, when a convolutional neural network is trained with a combination of images annotated manually and synthetically, there is a decreased need for proofreading because the reliability increases significantly. The researchers reported,

“By reducing the time needed for manual annotations and proofreading, our end-to-end pipeline achieves a 50-fold increase in throughput in comparison to full manual annotation for challenging brain imaging problems.”