Time-domain Full-Field OCT (FFOCT) acquires en-face images of a sample in a given depth. One of the biggest challenge to use FFOCT for in-vivo retinal imaging is the presence of retinal axial motion during image acquisition. To address this challenge, we previously proposed to couple a spectral-domain OCT (SD-OCT), where axial motion can be measured and later corrected by moving the FFOCT reference arm accordingly in a control loop fashion. However, due to the inherent temporal delay of the control-loop (typically 2-frame delay and 50Hz loop rate), the achieved precision was only around 10µm rms, against ideal 4µm rms (coherence gate of 8µm). Here, we propose to use learning-based prediction methods to enhance the precision of the axial retinal motion correction in real-time, thus improving the FFOCT robustness for in-vivo retinal imaging.
|