3D fast ray-traced distortion correction of indirectly viewed intraoperative retinal OCT for surgical tool path...

Intraoperative optical coherence tomography (iOCT) has emerged as a promising modality for real-time three-dimensional image guidance in retinal surgery, enabling visualization of different retinal layers and tool-tissue interactions with micrometer-scale precision. Indirect viewing systems, which employ additional optical elements to increase the surgical field of view (FOV) of the retina, are widely adopted in vitreoretinal procedures. However, the optical design of these indirect viewing systems introduces nonlinear optical distortions that compromise the spatial accuracy required for absolute positioning, such as might be required for robotic end effectors operating in the surgical FOV. In this paper, we present a computationally efficient method to achieve accurate distortion correction between the distorted optical coherence tomography (OCT) image space and the physical surgical workspace through a combination of optical ray tracing and parametric spline fitting. Our approach leverages physics-based ray tracing to characterize the complex optical transformations introduced by the indirect viewing system, then employs numerical spline interpolation to create a fast mapping function suitable for robotic control applications. The spline-based transformation achieves approximately 14 × faster computation than direct commercial software based optical ray tracing. Additionally, we developed a simplified porcine eye model encompassing the cornea through the crystalline lens for use in ray tracing simulations. Validation on ex vivo porcine eyes demonstrated that the proposed method reduced distortion errors along simulated surgical tool insertion paths. Across 30 OCT volumes from three eyes, the reconstructed length of a 7 mm straight tool path deviated from the open-path ground truth by 498 μm before correction and by 19 μm after correction. Over the same volumes, the correction reduced the residual registration error against the ground truth from 150 μm to 90 μm, and the straightness deviation of the reconstructed path from 226 μm to 137 μm, compared with 109 μm for the ground truth itself.