Practical aberration correction using deep transfer learning with limited experimental data.

Area of Science:

• Microscopy
• Optical Engineering
• Machine Learning

Background:

• Adaptive optics (AO) corrects aberrations to enhance image quality in microscopy.
• Traditional AO methods often rely on iterative aberration determination, which is time-consuming.
• Deep learning (DL) offers non-iterative aberration prediction but requires extensive training data.

Purpose of the Study:

• To address the data requirement challenge in DL for AO microscopy.
• To develop a practical DL approach for aberration prediction and correction using transfer learning.
• To validate the method's effectiveness with limited experimental data.

Main Methods:

• Employed transfer learning by pre-training a DL network on a large simulated dataset.
• Fine-tuned the pre-trained network using a small set of experimental data (24 samples).
• Extended aberration prediction to 25 Zernike modes and analyzed phase-diversity requirements.

Main Results:

• Achieved significant aberration reduction (average 73% decrease in RMS wavefront error) on experimental data for 10 Zernike modes.
• Demonstrated noticeable improvements with minimal fine-tuning data.
• Image capture and aberration inference rates are comparable to laser scanning microscopy acquisition times.

Conclusions:

• Transfer learning effectively overcomes the large dataset limitation for DL in AO microscopy.
• The proposed method offers a practical and efficient solution for aberration prediction and correction.
• The approach is compatible with iterative refinement for further performance enhancement.