Automated High-throughput Quantification Algorithm For Zebrafish Larvae Group

Over the last 40 years,the zebrafish(Danio rerio)has become an extremely attractive model organism for biomedical studies,including genetics,drug discovery,toxicology,developmental biology,behavior science,due to its numerous advantages,such as high similarity of gene and cardiovascular and nervous system to human,optical transparency,easiness of acquisition,and rapid developmental process.Among the experimental organisms of zebrafish,the larva has a unique value for its immaturity and motility.For one,larval motility,transparent body and developed organs illustrate more phenotypic information than cells and embryos;for another,larval body transparency and immature nervous system enable more studies and experiments that cannot be performed in the adult period.Due to these reasons,the zebrafish larva has gained widespread popularity.Video tracking has become a standard procedure for studying larval functional characteristics.Generally,scientists manipulate zebrafish larvae genetically or pharmaceutically and proceed to record the sequences of individual locomotive behaviors within a group for further analysis,including the study of larval movement characteristics,behavioral science investigations,and analysis of phenotypes resulting from genetic mutations,gene knockdown approaches,drugs and toxins.Typically,after this recording process,where large amounts of image raw data are generated,operators manually observe and label the image features like position and velocity of organisms one by one,frame by frame,which is a subjective and tedious task.As a result,an automatic and robust tracking algorithm is in massive demand.Despite numerous commercial tracking systems for model organisms,there is still a gap between the huge demand for larval monitoring systems and the availability of a reliable software for larval group behavior quantification,due to larval small size,transparent and similar appearance,frequent occlusions,and discontinuous kinematics.This paper proposes a powerful,robust,efficient,and high-throughput algorithm to quantify zebrafish larvae group movement.The key contributions of this work include: constructing a light convolution neural network and a topological analysis technique to visually track individual larvae in the case of intersections;designing an Adaptive Filter to tackle the discontinuous kinematics and obtain the optimal estimation of larval locomotive parameters;proposing an efficient data association protocol based on Intersection over Union(IOU)and Mahalanobis distance to link the associate objects.Experimental results demonstrate that this tracker achieved a multi-object tracking accuracy of more than 97% and a processing speed of over 30 frames per second.Its tracking precision is comparable to current state-of-the-art deep learning algorithms while achieving real-time tracking.