Exploring the use of deep learning models for accurate tracking of 3D zebrafish trajectories.

Abstract

Zebrafish are ideal model organisms for various fields of biological research, including genetics, neural transmission patterns, disease and drug testing, and heart disease studies, because of their unique ability to regenerate cardiac muscle. Tracking zebrafish trajectories is essential for understanding their behavior, physiological states, and disease associations. While 2D tracking methods are limited, 3D tracking provides more accurate descriptions of their movements, leading to a comprehensive understanding of their behavior. In this study, we used deep learning models to track the 3D movements of zebrafish. Videos were captured by two custom-made cameras, and 21,360 images were labeled for the dataset. The YOLOv7 model was trained using hyperparameter tuning, with the top- and side-view camera models trained using the v7x.pt and v7.pt weights, respectively, over 300 iterations with 10,680 data points each. The models achieved impressive results, with an accuracy of 98.7% and a recall of 98.1% based on the test set. The collected data were also used to generate dynamic 3D trajectories. Based on a test set with 3,632 3D coordinates, the final model detected 173.11% more coordinates than the initial model. Compared to the ground truth, the maximum and minimum errors decreased by 97.39% and 86.36%, respectively, and the average error decreased by 90.5%.This study presents a feasible 3D tracking method for zebrafish trajectories. The results can be used for further analysis of movement-related behavioral data, contributing to experimental research utilizing zebrafish.