Xu Guozheng, Smart Thomas J, Durech Eduard, Sarunic Marinko V
Department of Medical Physics and Biomedical Engineering, University College London, London WC1E 6BT, United Kingdom.
Institute of Ophthalmology, University College London, London WC1E 6BT, United Kingdom.
Biomed Opt Express. 2024 Jul 23;15(8):4795-4814. doi: 10.1364/BOE.528579. eCollection 2024 Aug 1.
Sensorless adaptive optics (SAO) has been widely used across diverse fields such as astronomy, microscopy, and ophthalmology. Recent advances have proved the feasibility of using the deep deterministic policy gradient (DDPG) for image metric-based SAO, achieving fast correction speeds compared to the coordinate search Zernike mode hill climbing (ZMHC) method. In this work, we present a multi-observation single-step DDPG (MOSS-DDPG) optimization framework for SAO on a confocal scanning laser ophthalmoscope (SLO) system with particular consideration for applications in preclinical retinal imaging. MOSS-DDPG optimizes target Zernike coefficients in a single-step manner based on 2 + 1 observations of the image sharpness metric values. Through simulations, MOSS-DDPG has demonstrated the capability to quickly achieve diffraction-limited resolution performance with long short-term memory (LSTM) network implementation. tests suggest that knowledge learned through simulation adapts swiftly to imperfections in the real system by transfer learning, exhibiting comparable performance to the ZMHC method with a greater than tenfold reduction in the required number of iterations.
无传感器自适应光学（SAO）已在天文学、显微镜学和眼科等多个领域得到广泛应用。最近的进展证明了将深度确定性策略梯度（DDPG）用于基于图像度量的SAO的可行性，与坐标搜索泽尼克模式爬山（ZMHC）方法相比，实现了更快的校正速度。在这项工作中，我们针对共焦扫描激光检眼镜（SLO）系统上的SAO提出了一种多观测单步DDPG（MOSS-DDPG）优化框架，特别考虑了其在临床前视网膜成像中的应用。MOSS-DDPG基于对图像清晰度度量值的2 + 1次观测，以单步方式优化目标泽尼克系数。通过仿真，MOSS-DDPG通过长短期记忆（LSTM）网络实现，已证明能够快速实现衍射极限分辨率性能。测试表明，通过仿真学到的知识通过迁移学习能迅速适应实际系统中的缺陷，其性能与ZMHC方法相当，所需迭代次数减少了十倍以上。
