A comparative study of deep learning-based zebrafish image segmentation methods

This study systematically evaluates several deep learning-based segmentation networks for zebrafish image analysis. With the same dataset and preprocessing procedures, we selected eleven representative segmentation models, including U-Net, SegNet, PSPNet, DeepLabv3+, Attention U-Net, HRNet, SegFormer, MASNet, the Segment Anything Model (SAM), PVT-EMCAD, and RWKV-UNet. We comprehensively compared their performance via various evaluation metrics, such as Dice coefficient, intersection-over-union, and mean pixel accuracy. Qualitative evaluations are combined with quantitative ones to compare and analyze the models' effectiveness in detecting intricate organ structures and phenotypic anomalies among zebrafish. The experimental results show that high-resolution maintenance, context-aggregation mechanisms, and attention mechanisms are the keys to improving the accuracy of small structure detection and boundary delineation, and the Transformer architecture also shows significant advantages in global dependency modeling. This study provides a general technique and theoretical support for high-throughput toxicological screening and morphological quantification of zebrafish images.