Single Scan OCT-based Retina Detection for Robot-assisted Retinal Vein Cannulation

Vitreoretinal surgery concerns a set of particularly demanding minimal invasive micro-surgical interventions at the retina. Micro-surgeons are targeting sub-millimeter-sized structures here. Tiny vessels or wafer-thin membranes are to be cannulated or need to be peeled off. The greatest care is to be displayed not to damage these fragile structures or to inadvertently injure the underlying retina. Damage to the latter is mostly irreparable and might cause permanent loss of vision. Despite the availability over excellent stereo microscopes, wide-angle lenses and powerful light source visualization remains a problem. Especially, the limited depth perception is still perceived as a major bottle-neck whereas efforts have been conducted to integrate sensing capability in today’s state-of-the-art instruments, so far, little effort has been paid to process the obtained sensor data and turns this into a reliable source of information upon which robot assistive guidance schemes could be endowed upon. This paper proposes a number of processing techniques tailored to Optical Coherence Tomography (OCT) measurements. The first results of the proposed algorithms show that it is feasible to extract good and reliable distance estimates from this otherwise rather noisy signal and from a fairly limited dataset. The used data are the so-called A-scans. These are OCT measurements consisting of a single-line image that could be captured by an instrument-mounted fiber through which the OCT signal passes back-and-forth. However, in this work, we perform a pilot study whereby the employed A-scans are extracted from B-scans that are captured by a microscope-mounted OCT scanner, rather than obtained from a probe. The performance of a first embodiment of the algorithm that is based on an Unscented Kalman Filter (UKF) is compared to the performance of a second embodiment that relies on a Particle Filter (PF), focusing on the issues in filter initialization and the tracking quality. Finally, results of UKF and PF executions on a validation dataset are presented. Read More: https://www.worldscientific.com/doi/abs/10.1142/S2424905X18400056