Deep Convolutional Neural Networks for Autofocus Control on a C. elegans Tracking System.

Area of Science:

• Microscopy and Imaging Technologies
• Computational Biology
• Machine Learning in Biosciences

Background:

• Accurate focal positioning is critical for high-resolution microscopy of dynamic biological samples, such as Caenorhabditis elegans.
• Existing focusing methods often lack the speed required for real-time tracking of moving specimens.

Purpose of the Study:

• To develop a rapid, one-shot focusing method using convolutional neural networks (CNNs) for live microscopy.
• To enhance focusing accuracy and efficiency, eliminating the need for iterative optical axis scanning.

Main Methods:

• A novel CNN-based approach for direct prediction of optimal focusing distance.
• Implementation of a new data augmentation technique to improve model performance without additional data.
• Training and comparison of various CNN architectures, including ConvNext V2, on z-stacks of images.

Main Results:

• The proposed one-shot CNN method significantly accelerates focal positioning compared to traditional scanning techniques.
• A new data augmentation strategy was validated to enhance model accuracy without requiring more experimental data.
• ConvNext V2 demonstrated superior performance over other evaluated architectures for this focusing task.

Conclusions:

• The developed CNN-based method offers a fast and accurate solution for real-time focusing in live microscopy.
• The novel data augmentation technique effectively improves model generalization and performance.
• ConvNext V2 shows promise as a leading architecture for automated microscopy focusing applications.