Deep Reinforcement Learning for Wavefront Sensorless Adaptive Optics

Many retinal diseases can be diagnosed using retinal imaging techniques such as Optical Coherence Tomography (OCT) and Scanning Laser Ophthalmoscopy (SLO). To obtain clear images, optical distortions are corrected by using Adaptive Optics (AO) to adjust the wavefront before it gets to the image sensor. Wavefront Sensorless Adaptive Optics (SAO) is becoming increasing used with advances in computational power, however significant acquisition time is a concern. In this thesis, deep reinforcement learning (DRL) is applied to SAO to improve on existing SAO methods for up to 18 Zernike modes by reducing the number of acquisitions required. A range of single and multi-agent architectures using DRL were explored and implemented, trained, and tested in silico. The final system used a Long Short-Term Memory (LSTM) architecture with two agents and outperformed Zernike Mode Hill Climbing (ZMHC) in multiple scenarios.