Deep reinforcement learning for automatic defocus correction using OCT image intensity

Optical coherence tomography (OCT) image stability often suffers during in vivo imaging of the retina due to axial motion of the subject's head and changes in their visual focus. Ocular accommodation can actively adjust the focus, affecting the axial intensity distribution across the retinal cross-section and the lateral resolution of the target layers. Axial motion shifts the retinal image and affects en face visualization of retinal layers. We present an automated procedure for stabilization of axial motion and focus during OCT retinal image acquisition using deep reinforcement learning (DRL) for defocus correction. The correction process requires only B-scan images as inputs, making it suitable for real-time correction. In silico training and in vivo fine-tuning experiments have been conducted and presented to validate the performance of the correction procedure for retinal imaging.