A YOLO based software for automated detection and analysis of rodent behaviour in the open field arena

• Published in: Computers in biology and medicine 2021-07, Vol.134, p.104474-104474, Article 104474
• Main Authors: Hatton-Jones, Kyle M., Christie, Corey, Griffith, Tia A., Smith, Amanda G., Naghipour, Saba, Robertson, Kai, Russell, Jake S., Peart, Jason N., Headrick, John P., Cox, Amanda J., du Toit, Eugene F.
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Animal behavior; Animal models; Automated; Automation; Behavior; Behavioural tests; Cardiovascular disease; Deep learning; Exploratory behavior; Field tests; Investigations; Learning algorithms; Locomotion; Locomotor activity; Machine learning; Neurosciences; Open-field behavior; Open-field test; Reproducibility; Rodents; Software; Spatial analysis; Validation studies; YOLO
• ISSN: 0010-4825; 1879-0534
• DOI: 10.1016/j.compbiomed.2021.104474

Rodent models are important in mechanistic studies of the physiological and pathophysiological determinants of behaviour. The Open Field Test (OFT) is one of the most commonly utilised tests to assess rodent behaviour in a novel open environment. The key variables assessed in an OFT are general locomotor activity and exploratory behaviours and can be assessed manually or by automated systems. Although several automated systems exist, they are often expensive, difficult to use, or limited in the type of video that can be analysed. Here we describe a machine-learning algorithm – dubbed Cosevare – that uses a trained YOLOv3 DNN to identify and track movement of mice in the open-field arena. We validated Cosevare's capacity to accurately track locomotive and exploratory behaviour in 10 videos, comparing outputs generated by Cosevare with analysis by 5 manual scorers. Behavioural differences between control mice and those with diet-induced obesity (DIO) were also documented. We found the YOLOv3 based tracker to be accurate at identifying and tracking the mice within the open-field arena and in instances with variable backgrounds. Additionally, kinematic and spatial-based analysis demonstrated highly consistent scoring of locomotion, centre square duration (CSD) and entries (CSE) between Cosevare and manual scorers. Automated analysis was also able to distinguish behavioural differences between healthy control and DIO mice. The study found that a YOLOv3 based tracker is able to easily track mouse behaviour in the open field arena and supports machine learning as a potential future alternative for the assessment of animal behaviour in a wide range of species in differing environments and behavioural tests. •Freely available machine-learning algorithms were used to create a tracking software for rodent behaviour.•The software was successful in tracking in the open-field test and determining exploratory behaviour.•Behavioural differences were observed in a mouse model of obesity.