Learning to autofocus in whole slide imaging via physics-guided deep cascade networks

Area of Science:

• Digital Pathology
• Computational Imaging
• Machine Learning in Microscopy

Background:

• Whole slide imaging (WSI) is crucial for digital pathology, but its performance relies heavily on accurate autofocusing.
• Conventional autofocusing techniques are often slow or require hardware modifications, limiting their integration with existing WSI systems.

Purpose of the Study:

• To develop an effective, learning-based autofocusing method for WSI that is fast and requires no hardware changes.
• To address the limitations of current autofocusing methods in terms of speed and compatibility.

Main Methods:

• A novel deep cascade network approach inspired by the asymmetry of optical aberrations in WSI images.
• A binary classification network to distinguish positive/negative defocus based on image characteristics.
• A refocusing network to learn the mapping between defocus images and defocus distance.

Main Results:

• The proposed method achieves accurate and high-speed autofocusing without optical hardware modifications.
• The deep cascade networks effectively handle optical aberrations for improved autofocusing quality.
• Experimental results show superior performance compared to existing autofocusing methods.

Conclusions:

• The learning-based autofocusing method offers a significant advancement for WSI systems.
• This approach enhances the efficiency and accuracy of digital pathology workflows.
• The method provides a practical solution for improving WSI performance through advanced image analysis.