Longitudinal Identification of Zebrafish Individuals by Deep Learning

Abstract

The zebrafish (Danio rerio) is a critical vertebrate model organism in biomedical research. Accurate identification and longitudinal tracking of individual fish within cohorts enable linking genotype to complex phenotypes, essential for elucidating the molecular mechanisms underlying disease pathogenesis and trait formation. Nevertheless, the establishment of reliable long-term individual identification remains a significant challenge, primarily due to their diminutive size and minimal interindividual morphological variations. We developed ESC-IDNet, a dual-stage deep learning cascade architecture deployed on the FishIndivID platform (http://bioinfo.ihb.ac.cn/fishindivid), enabling high-precision identification. Trained and tested on [~]300,000 images from 450 zebrafish (31-122 dpf), our lateral body-based identification maintained >95% accuracy with updates only every 20 days, significantly outperforming dorsal head identification (requiring [~]13-day updates). Both views achieved 100% accuracy in 1-2 day tasks. ESC-IDNet surpassed alternatives in segmentation, alignment, feature extraction, and feature matching. This framework provides a robust, transferable paradigm for individual identification in fish species.