Transform- and multi-domain deep learning for single-frame rapid autofocusing in whole slide imaging

Area of Science:

• Digital pathology
• Microscopy imaging
• Computational imaging

Background:

• Whole slide imaging (WSI) systems are crucial for digital pathology, with autofocusing performance directly impacting image quality and throughput.
• Traditional autofocusing methods involve axial scanning and merit function optimization, which can be time-consuming and limit system speed.

Purpose of the Study:

• To develop and evaluate a deep learning-based autofocusing method for WSI that eliminates the need for axial scanning.
• To investigate the effectiveness of using transform domain features (Fourier spectrum, autocorrelation) for improved autofocusing accuracy and robustness.

Main Methods:

• Utilized deep convolutional neural networks (CNNs) trained on approximately 130,000 images with varying defocus distances.
• Explored spatial image information, Fourier spectrum, and autocorrelation of images as input features for the CNNs.
• Evaluated autofocusing performance under different illumination settings: incoherent Kohler, partially coherent (two plane waves), and one plane wave.

Main Results:

• Deep learning models incorporating transform domain features significantly outperformed models relying solely on spatial information.
• Achieved a focusing error of approximately 0.5 µm, well within the 0.8 µm depth-of-field range.
• The developed method demonstrated robustness across different illumination conditions and requires minimal hardware modification.

Conclusions:

• Deep learning-based autofocusing using transform domain features offers a highly accurate and efficient solution for WSI systems.
• This approach enables on-the-fly image capture without focus map surveying, enhancing WSI and time-lapse microscopy applications.
• The open-source dataset and methodology provide valuable resources for advancing deep learning in microscopy.