Image metric-based multi-observation single-step deep deterministic policy gradient for sensorless adaptive optics

Area of Science:

• Biomedical optics
• Computational imaging
• Ophthalmology

Background:

• Sensorless adaptive optics (SAO) is crucial for improving image quality in various imaging modalities.
• Deep deterministic policy gradient (DDPG) has shown promise for faster SAO compared to Zernike mode hill climbing (ZMHC).
• Preclinical retinal imaging requires precise aberration correction for high-resolution visualization.

Purpose of the Study:

• To introduce a multi-observation single-step DDPG (MOSS-DDPG) optimization framework for SAO.
• To apply MOSS-DDPG to a confocal scanning laser ophthalmoscope (SLO) for preclinical retinal imaging.
• To evaluate the performance of MOSS-DDPG in terms of speed and accuracy.

Main Methods:

• Developed a MOSS-DDPG framework optimizing N Zernike coefficients using 2N+1 image sharpness metric observations.
• Implemented MOSS-DDPG with a long short-term memory (LSTM) network.
• Conducted in silico simulations and in situ tests on a confocal SLO system.

Main Results:

• MOSS-DDPG achieved diffraction-limited resolution in simulations.
• Transfer learning enabled rapid adaptation of simulation-learned knowledge to real-world system imperfections.
• In situ tests showed comparable performance to ZMHC with over a tenfold reduction in iterations.

Conclusions:

• MOSS-DDPG offers a highly efficient approach for SAO in preclinical retinal imaging.
• The framework demonstrates rapid convergence and robust performance in real-world conditions.
• MOSS-DDPG represents a significant advancement for high-resolution retinal imaging applications.