Deep learning for robotic-assisted OCT

Robotics has been leveraged in OCT to extend the scale of imaging. It still has two technical issues should be addressed before practical applications: (1) For the probe-to-surface control, the rule-based feature extraction is susceptible to imaging artifacts of OCT. (2) The use of ultra-fast swept-sources improve the speed of scanning, but leads to a significant increase in system cost. Here we attempt to address the above issues with the automatic feature extraction and compressive sensing capabilities of deep learning. We found that a trained U-shape segmentation network could implicitly and accurately detect the surfaces of imaging objects under the influences of various OCT artifacts. For the second issue, we use sampling densities lower than Nyquist sampling theorem for data acquisition, and then reconstruct it in real-time at a high sampling density using a trained neural network. It should be noted that the end-to-end, fast inference of trained deep learning models would not degrade the acquisition speed of robotic OCT. We think the adoption of our methods may promote more medical and industrial applications for the robotic-assisted OCT.