Deep-Pose-Tracker: a unified model for behavioural studies of Caenorhabditis elegans

Tracking and analyzing animal behaviour is a crucial step in fields such as neuro-science and developmental biology. Behavioral studies in the nematode C. elegans, for example, help in understanding how organisms respond to external cues and how the specific physiological responses link to either instantaneous or learned behavior. Although tracking behaviour through locomotion patterns and posture dynamics are routine, they become laborious, time-consuming tasks when performed manually. Automation of this process is therefore crucial for accurate and fast detection and analysis. To this end, in this work, we report the development of Deep Pose Tracker (DPT), an end-to-end deep learning model designed to automate the analysis of posture dynamics and locomotion behaviour of C. elegans. This YOLO (You Only Look Once)-based model enables automatic detection and tracking of these worms while measuring essential behavioural features like motion speed, orientation, forward or reverse locomotion and complex body bends like omega turns. In addition, it includes eigenworm decomposition in order to analyze different body shapes that these tiny worms make during their motion, and represent the overall posture dynamics in a low-dimensional space. Our DPT model can generate highly accurate data, with very high inference speed while being user-friendly. DPT, therefore, can be a valuable toolkit for researchers studying behaviour under different environmental stimuli.