Deep learning-based adaptive optics for light sheet fluorescence microscopy.

Area of Science:

• Biomedical imaging
• Optical microscopy
• Computational imaging

Background:

• Light sheet fluorescence microscopy (LSFM) enables high-speed imaging of cleared tissues at cellular resolution.
• LSFM performance is degraded by sample-induced optical aberrations, particularly in deep tissue imaging.
• Current sensorless adaptive optics methods for aberration correction are slow, requiring thousands of images.

Purpose of the Study:

• To develop a fast and accurate method for estimating and correcting optical aberrations in LSFM.
• To improve image quality and enable deeper imaging of cleared biological specimens.
• To address the limitations of existing adaptive optics techniques in high-throughput microscopy.

Main Methods:

• Utilized deep learning algorithms to estimate sample-induced aberrations from minimal image data (two images per region).
• Implemented a deformable mirror for aberration correction based on deep learning estimations.
• Introduced a novel sampling strategy for efficient neural network training.
• Compared two distinct deep learning network architectures for aberration estimation.

Main Results:

• Deep learning accurately estimated aberrations from only two images, significantly reducing acquisition time.
• Aberration correction using a deformable mirror demonstrably improved image quality in cleared tissues.
• The developed method offers a substantial speed improvement over conventional sensorless adaptive optics.
• Both tested network architectures provided effective aberration correction.

Conclusions:

• Deep learning provides an efficient and rapid solution for correcting optical aberrations in LSFM.
• This approach enhances image quality and facilitates deeper, more reliable imaging of cleared tissues.
• The method overcomes critical speed limitations of traditional adaptive optics, enabling faster and more comprehensive biological sample analysis.