Cross-Shaped Heat Tensor Network for Morphometric Analysis Using Zebrafish Larvae Feature Keypoints.

Area of Science:

• Developmental biology
• Bioinformatics
• Machine learning

Background:

• Deep learning-based morphometric analysis of zebrafish is crucial for non-destructive disease diagnosis.
• Challenges exist in obtaining clear organ boundaries from external observation of zebrafish larvae.

Purpose of the Study:

• To develop a deep learning method for precise endpoint detection of zebrafish organ features.
• To enable quantitative morphometric analysis and phenotype extraction in zebrafish larvae.

Main Methods:

• Proposed the cross-shaped heat tensor network (CSHT-Net) for feature point detection.
• Introduced a novel keypoint training method (cross-shaped heat tensor) and feature extractor (combinatorial convolutional block).
• Utilized a dataset of 4389 zebrafish larvae micrographs (120 h post-fertilization).

Main Results:

• CSHT-Net achieved an average precision (AP) of 83.2% and average recall (AR) of 85.8%.
• The model outperforms existing keypoint detection techniques.
• Demonstrated enhanced ability to learn continuous, strip-like features compared to heatmap-based methods.

Conclusions:

• The CSHT-Net framework enables robust phenotype extraction and reliable morphometric analysis in zebrafish larvae.
• This approach facilitates efficient hazard identification for chemicals and medical products.
• Simplifying organ areas to polygons and focusing on endpoint positioning addresses challenges in external observation.