Towards an Automated Zebrafish-based Toxicity Test Model Using Machine Learning

Abstract Zebrafish animal is considered as one of the most suitable animals to test toxicity of compounds due many features such as transparency and a large number of embryos produced in each mating. The main problem of the zebrafish-based toxicity test is the manual inspection of thousands of animals images in different phases and this is not feasible enough for the analysis, i.e. it is slow and may be inaccurate process. To help addressing this problem, in this paper, an automated classification of alive (healthy) and coagulant (died because of toxic compounds) zebrafish embryos are proposed. The embryos’ images are used to extract some features using the Segmentation-based Fractal Texture Analysis (SFTA) technique. The Rotation Forest classifier is then used to match between testing and training features (i.e. to classify alive and coagulant embryos). The experiments have proved that choosing threshold value of SFTA technique and the size of the rotation forest classifier have a great impact on the classification accuracy. With accuracy around 99.98%, the experimental results have showed that the proposed model is a very promising step toward a fully automated toxicity test during drug discovery.