Fourier-based structural metric with a CNN-assisted focus decision for robust autofocusing in microscopy.

Area of Science:

• Microscopy and Imaging Science
• Computational Imaging
• Materials Science

Background:

• Autofocusing in microscopy is crucial for image quality but challenging with varying illumination and sample types.
• Conventional methods like gradient-based measures struggle with noise, saturation, and transparency.
• Polynomial fitting methods exhibit limitations such as boundary oscillations and loss of sensitivity.

Purpose of the Study:

• To develop a robust and generalizable autofocus method for microscopy.
• To overcome the limitations of existing autofocusing techniques.
• To improve focal accuracy across diverse sample types and imaging conditions.

Main Methods:

• A novel Fourier-based structural metric (f_FS(z)) was developed to stably preserve structural peaks under various illumination levels.
• A convolutional neural network (CNN) was employed for assisted focus decision-making.
• The method integrates Fourier-domain analysis with CNN-based region discrimination (higher-plane, lower-plane, transparent) for refined focal positioning.

Main Results:

• The f_FS(z) metric demonstrated distinguishable patterns (local peaks for low/high intensity, global peaks for overexposure), simplifying focus signatures.
• The CNN effectively discriminated between different sample regions, enabling precise focal point selection.
• Validation on metallic alloys and biological tissues showed consistent and accurate focal positioning, surpassing conventional methods.

Conclusions:

• The combined Fourier analysis and CNN approach provides a robust and generalizable autofocus solution for microscopy.
• This method significantly improves focal accuracy, especially in challenging imaging scenarios where traditional methods fail.
• The technique is applicable to a wide range of samples, including reflective, textured, and biological tissues.