Magnification-Invariant Retinal Distance Estimation for Vitreoretinal Surgery Using a Laser Aiming Beam

Abstract

Background Virtual fixtures in robot-assisted retinal surgery require knowledge of the position of the retina with respect to the surgical tool to be effective. Retinal surface estimation is a difficult problem due to the lack of features in the captured microscope imagery and a complex light path. A laser aiming beam attached to the tool can be easily detected in the microscope imagery and provide valuable information about the location of the surface.

Methods We propose using the area of a laser aiming beam attached to the surgical tool to determine the distance of the tool from the retina. This area was modified in accordance with the tool width to ensure independence from microscope magnification. Retinal distance is predicted using a dual Kalman filter that combines distance inferred from this metric with information from an optical tracker that tracks the position of the tool in a global coordinate system. This updates both the state and parameters of the system in parallel and allows us to predict retinal distance even with errors in initial parameters.

Results The laser metric's independence from microscope magnification is demonstrated by plotting the metric at 3 different magnifications for a number of angles. We also predict the distance of the tool from the retina for various random angles at each magnification with median errors of less than 100 μm. Finally, we predict distance at each magnification during freehand motion and validate our results using a force sensor placed underneath the phantom.

Conclusions Using the area of laser aiming beam attached to the surgical tool, our method can predict the distance of the surgical tool to the retina with errors that are acceptable for implementing virtual fixtures during robot-assisted retinal surgery. The predicted distance is also independent of microscope magnification and can work when initial parameters are not precisely known. Future work will involve adapting this method to in vivo environments and further reduction in prediction errors.