Deep Learning Based Real-Time OCT Image Segmentation and Correction for Robotic Needle Insertion Systems

Abstract

This article proposes deep learning based real-time optical coherence tomography (OCT) image segmentation and correction algorithm for vision-based robotic needle insertion systems that can be used in DALK (deep anterior lamellar keratoplasty) surgery. The proposed algorithm provides the position of the needle tip, the lower boundary of the tissue, and the marginal insertion depth solving traditional issues of OCT images like refractive error, optical noise from surgical tools, and the slow speed of volumetric scanning. Through the ex-vivo experiment using 10 porcine corneas, the performance of the proposed algorithm with a robotic system was verified. The segmentation errors were 7.4 mu m for the upper boundary, 10.5 mu m for the lower boundary, and 3.6 mu m for the needle tip. The difference in needle slope between the outside and inside of the cornea was dramatically reduced from 5.87 degree to 0.78 degree. The frame rate of the OCT image was 9.7 Hz, and the time delay of the image processing algorithm was 542.6 ms for 10 images of 512 x 512 pixels. The results of the proposed algorithm were compared with those of the previous studies.