Deep reinforcement learning for automatic defocus correction using OCT image intensity

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) imaging of the retina is susceptible to instability from subject motion and focus changes.
• Axial motion and ocular accommodation-induced defocus degrade image quality, affecting visualization of retinal layers and lateral resolution.
• Current methods may not adequately address both axial motion and focus-related artifacts in real-time OCT retinal imaging.

Purpose of the Study:

• To develop and validate an automated procedure for stabilizing axial motion and focus in OCT retinal imaging.
• To implement a deep reinforcement learning (DRL) approach for real-time defocus correction using B-scan images.
• To improve the clarity and reliability of OCT retinal imaging for diagnostic purposes.

Main Methods:

• An automated procedure utilizing deep reinforcement learning (DRL) was developed for simultaneous stabilization of axial motion and focus.
• The DRL model was trained using in silico data and fine-tuned with in vivo experiments.
• The correction method requires only B-scan images as input, enabling real-time application.

Main Results:

• The DRL-based procedure effectively stabilized OCT retinal images against axial motion and defocus.
• Real-time correction was achieved, significantly improving the quality of retinal cross-sectional and en face visualizations.
• Validation through in silico and in vivo experiments confirmed the procedure's performance.

Conclusions:

• The presented automated DRL procedure offers a robust solution for enhancing OCT retinal image stability.
• This method holds potential for improving diagnostic accuracy in ophthalmology by providing clearer retinal images.
• The real-time, B-scan-based correction is suitable for clinical application in OCT retinal imaging.