A Deep Learning Approach to Measure Visual Function in Zebrafish

Visual behaviour in zebrafish, often measured by the optokinetic reflex (OKR), serves as a valuable model for studying aspects of human neurological and ocular diseases and for conducting therapeutic or toxicology assays. Traditional methods for OKR analysis often rely on binarization techniques (threshold-based conversion of images to black and white) or costly software, which limits their utility in low-contrast settings or hypopigmented disease models. Here, we present a novel deep learning pipeline for OKR analysis, using ResNet-50 within the DeepLabCut framework in a Python Version 3.10 environment. Our approach employs object tracking to enable robust eye movement quantification, regardless of variations in contrast or pigmentation. OKR responses were elicited in both wild-type and slc45a2 (albino) mutant zebrafish larvae at 5 days post-fertilisation, using a mini-LED arena with a rotating visual stimulus. Eye movements were recorded and analysed using both conventional software and our deep learning approach. We demonstrate that the deep learning model achieves comparable accuracy to traditional methods, with the added benefits of applicability in diverse lighting conditions and in hypopigmented larvae. Statistical analyses, including Bland-Altman tests, confirmed the reliability of the deep learning model. While this study focuses on 5-day-old zebrafish larvae under controlled conditions, the pipeline is adaptable across developmental stages, pigmentation types, and behavioural assays. With appropriate adjustments to experimental parameters, it could be applied to broader behavioural studies, including social interactions and predator-prey dynamics in ocular and neurological disease models.