New deep learning model uses video to measure embryo development

Research led by the University of Plymouth has shown that a new deep learning AI model can identify from video what happens during embryo development and when.

Published in Journal of Experimental BiologyThe study, titled “Dev-ResNet: Automatic Developmental Event Detection Using Deep Learning,” highlights how a model called Dev-ResNet can identify the occurrence of key functional developmental events in pond snails, such as cardiac function, crawling, hatching, and even death.

A key innovation in this research is the use of 3D models that exploit the changes that occur between frames of video, as opposed to the traditional use of still images, and allow the AI ​​to learn from these features.

Using video, Dev-ResNet reliably detected features ranging from the first heartbeat and crawling behaviour to shell formation and hatching, revealing previously unknown sensitivity of a range of features to temperature.

Although the study used snail embryos, the authors say the model is broadly applicable to all species and provide comprehensive scripts and documentation for applying Dev-ResNet to a range of biological systems.

In the future, the technology could help accelerate our understanding of how climate change and other external factors affect humans and animals.

The research was led by Ziad Ibbini, a PhD candidate who studied a BSc Conservation Biology at university before taking a year off to hone his software development skills and then proceeding with his PhD. He designed, trained and tested Dev-ResNet himself.

“Delineating developmental events – figuring out what happens early in an animal's development – is very challenging, but extremely important because it helps us understand variations in the timing of events across species and environments,” he said.

“Dev-ResNet is a small and efficient 3D convolutional neural network that can detect developmental events using video, and is relatively easy to train on consumer hardware.

“The only real limitation is in creating the data to train the deep learning model. We know the model works; we just need to provide it with the right training data.”

“We want to provide the broader scientific community with tools that can better understand how different factors affect species development and identify ways to protect them. We see Dev-ResNet as a big step in that direction.”

Dr Oli Tills, lead author of the paper and UKRI Future Leaders Research Fellow, added: “This research is important on a technical level, but also because it advances our understanding of biological development – something that the Ecophysiology and Development Research Group at the University of Plymouth has been investigating for over 20 years.”

“This groundbreaking achievement would not have been possible without deep learning, and it is exciting to think where this new capability will take us in studying animals during one of the most dynamic periods of their lives.”