4D microscope-integrated optical coherence tomography has emerged as a valuable tool for ophthalmic surgery. However, it is still not clear how to make the best, computationally efficient use of the abundant data generated. Many techniques require accurate registration of successive volume streams. To address this need, we present a real-time machine learning method for lateral registration based on en face projections. Our proposed network predicts the lateral translational and rotational offset, which can be corrected. As ground truth, we use homography matrices from the displacements of en face images. Our pipeline thus allows for real-time, rigid registration of en face maps derived from sequentially acquired volumes, and is able to accurately correct for lateral distortions at a volume rate of 20 Hz. We therefore believe that this method could enable real-time rigid volume registration for 4D microscopy integrated optical coherence tomography.
|