Zebrafish identification with deep CNN and ViT architectures using a rolling training window.

Area of Science:

• Animal Behavior and Physiology
• Machine Learning in Biology
• Bioinformatics

Background:

• Individual tracking of zebrafish in laboratory settings presents significant challenges due to complex and dynamic environmental conditions.
• Existing methods for zebrafish identification are often invasive or lack robustness over extended periods.
• Machine learning, especially neural networks, offers potential for developing adaptive and non-invasive identification systems.

Purpose of the Study:

• To develop and evaluate a novel rolling window training technique for robustly identifying individual zebrafish.
• To assess the efficacy of this technique with convolutional neural networks (CNNs) and vision transformers (ViTs).
• To analyze the contribution of visual features (shape, pattern, color) to identification accuracy.

Main Methods:

• Implementation of a rolling window training approach for CNN and ViT models.
• Utilizing open-source machine learning architectures for zebrafish identification.
• Systematic modification of training images to analyze the impact of shape, pattern, and color.
• Comparison of performance against other prevalent machine learning models.

Main Results:

• The rolling window training technique demonstrated high-fidelity identification of individual maturing zebrafish over several weeks.
• This method significantly reduces the need for continuous retraining with new image datasets for both CNN and ViT models.
• Analysis revealed the distinct contributions of shape, pattern, and color to the CNN classifier's success.

Conclusions:

• The rolling window training technique offers a promising solution for robust, long-term individual zebrafish tracking.
• This approach enhances the adaptability of machine learning models to changing conditions in biological studies.
• Understanding feature importance aids in optimizing future real-time zebrafish identification systems.