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

Area of Science:

• * Computational Biology
• * Image Analysis
• * Deep Learning

Background:

• * Zebrafish are crucial models in toxicology and developmental biology.
• * Accurate image segmentation is vital for high-throughput analysis.
• * Existing deep learning models require systematic evaluation for this specific application.

Purpose of the Study:

• * To systematically evaluate eleven deep learning segmentation networks for zebrafish image analysis.
• * To compare model performance using quantitative and qualitative metrics.
• * To identify key architectural features for accurate segmentation of zebrafish structures.

Main Methods:

• * Eleven deep learning segmentation models (U-Net, SegNet, PSPNet, DeepLabv3+, Attention U-Net, HRNet, SegFormer, MASNet, SAM, PVT-EMCAD, RWKV-UNet) were applied to the same zebrafish dataset.
• * Standardized preprocessing and evaluation metrics (Dice, IoU, mean pixel accuracy) were used.
• * Both quantitative and qualitative assessments were performed.

Main Results:

• * Models incorporating high-resolution maintenance, context aggregation, and attention mechanisms showed superior performance.
• * Transformer-based architectures demonstrated advantages in modeling global dependencies.
• * Specific models excelled in detecting small structures and delineating boundaries accurately.

Conclusions:

• * Key deep learning components for effective zebrafish image segmentation identified.
• * Findings support the use of advanced deep learning for automated toxicological screening and morphological quantification.
• * The study provides a framework for selecting optimal models for zebrafish image analysis.