A deep-learning-based workflow to deal with the defocusing problem in high-throughput experiments.

Area of Science:

• Biomaterials Science
• Microscopy
• Computational Biology

Background:

• High-throughput biomaterials research necessitates automated techniques for efficient data acquisition.
• Microscopy is crucial in biomaterials science, but defocused images can arise from experimental noise and mechanical errors.
• Existing automatic focusing methods are not always sufficient to handle all defocusing scenarios.

Purpose of the Study:

• To develop a deep-learning-based method for automatic sorting and reconstruction of defocused cell images.
• To address the challenge of image defocusing in high-throughput cell-based experiments.
• To improve the quality and utility of microscopy data in biomaterials research.

Main Methods:

• A comprehensive dataset of phase-contrast cell images under varied conditions was prepared.
• A deep learning model was trained for automatic sorting and reconstruction of defocused images.
• A separate architecture was employed for subcellular-level reconstruction of heavily defocused images.

Main Results:

• The deep-learning method achieved high accuracy (over 0.993) with a simple network architecture.
• The training time was less than half that of the classical ResNetV2 architecture.
• Subcellular-level reconstruction of heavily defocused images was successfully demonstrated.

Conclusions:

• The developed deep-learning workflow effectively sorts and reconstructs defocused cell images.
• The method requires no a priori knowledge of focusing algorithms, offering broad applicability.
• This intelligent workflow is valuable for high-throughput and time-lapse cell imaging in biomaterials research.