Automated staging of zebrafish embryos with deep learning.

Area of Science:

• Developmental biology
• Bioimage analysis
• Machine learning in science

Background:

• Zebrafish (Danio rerio) are crucial biomedical models.
• Detecting developmental delays in zebrafish embryos is vital but often manual, subjective, and time-consuming.
• Automated methods are needed to improve the efficiency and objectivity of developmental delay assessment.

Purpose of the Study:

• To develop and validate KimmelNet, a deep learning model for predicting zebrafish embryo age from 2D brightfield images.
• To enable high-confidence detection and quantification of developmental delays in zebrafish embryo populations.
• To provide a scalable and objective tool for developmental biology research.

Main Methods:

• A deep learning model (KimmelNet) was trained to predict embryo age in hours post-fertilisation from 2D brightfield microscopy images.
• The model's predictions were compared against established zebrafish staging methods.
• Transfer learning was employed to enhance model generalisation to new datasets.

Main Results:

• KimmelNet accurately predicts zebrafish embryo age, showing close agreement with established staging.
• The model reliably detects population-level developmental delays with high confidence using minimal data (as few as 100 images).
• KimmelNet demonstrates generalisation capabilities on unseen data, with performance boosted by transfer learning.

Conclusions:

• KimmelNet offers an automated, accurate, and efficient method for analysing zebrafish embryo development.
• This deep learning approach significantly improves upon manual observation for detecting developmental delays.
• The KimmelNet framework is adaptable for other model organisms, with broad potential applications in developmental biology.