Towards Safer Retinal Surgery through Chance Constraint Optimization and Real-Time Geometry Estimation

Abstract

Safely navigating a surgical tool to a desired location on the surface of the retina during retinal surgery relies on extreme precision and surgical skills. Damage to the delicate retinal tissue often occurs. Previous work demonstrated an approach for robot-assisted navigation in eye surgery using imitation learning and optimal control. To further enhance safety, we present a framework that combines real-time eye geometry estimation and chance-constrained optimal control to bound the probability for tissue damage during autonomous robotic navigation. A neural network is trained to predict the relative location of 3-D points on the retina with respect to the current tool-tip position through expert demonstrations. During inference, a local geometry of the retina is estimated using weighted least squares formulation based on these learned predictions and then employed as a probabilistic collision chance constraint in an optimal control framework. The proposed approach is experimentally validated using a phantom silicone eye suitable for vein cannulation testing. We statistically demonstrate that the network predictions become more accurate as the surgical tool approaches closer to the retinal surface and that the measured mean errors along the lateral directions for navigating to a given point or for a vessel-following task are below 0.1 mm. These results indicate that the proposed technique could serve as a basis to further develop robot-assisted retinal microsurgery with enhanced safety.