Establishing a reference focal plane using convolutional neural networks and beads for brightfield imaging

Area of Science:

• Microscopy and Image Analysis
• Computational Biology
• Biomedical Engineering

Background:

• Microscopy measurements suffer from repeatability issues due to sample heterogeneity, stage positioning, and slide thickness.
• Existing methods for establishing focal planes lack generalizability across different instruments.

Purpose of the Study:

• To develop a generalizable and highly accurate method for defining a reference effective focal plane (REFP).
• To improve the repeatability of image analytics measurements in microscopy.

Main Methods:

• Refined REFP definition using cubic spline fitting and cross-experiment data sharing.
• Implemented a convolutional regression neural network (ResNet 18) trained on control bead images.
• Validated the method on two different optical systems.

Main Results:

• The ResNet 18 network predicts REFP location with high accuracy (within 7.5 µm uncertainty).
• The method demonstrates generalizability across multiple microscopes and experimental conditions.
• Achieved high accuracy using as few as 6 beads per image.

Conclusions:

• The proposed method enhances measurement repeatability and accuracy in microscopy.
• The use of control beads and deep learning offers a robust solution for focal plane determination.
• This approach is applicable to diverse optical systems, advancing quantitative microscopy.