Free papers detail | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
Operational digital twins for embodied ophthalmic clinic

Authors
Xinyuan Wu, Mingguang He, Danli Shi

Presenting
Xinyuan Wu

PURPOSE:
To investigate how operational digital twins can be constructed to support embodied AI development in outpatient ophthalmic clinical workflows.

METHODS:
Ophthalmic clinic scenes were reconstructed using a single-image 3D Gaussian Splatting pipeline. Visual reconstruction quality was evaluated with structural similarity index measure (SSIM), learned perceptual image patch similarity (LPIPS), and geometry-related measures. Image-space editing was applied for scene reconfiguration and evaluated by masked and non-masked SSIM to quantify local modification and global preservation. Representative scenes were integrated with operational meshes, and preliminary interaction readiness was assessed in a representative task using touch-point repeatability, point-to-plane residuals, and directional dispersion relative to the fitted interaction plane.

METHODS:

Ophthalmic clinic scenes were reconstructed using a single-image 3D Gaussian Splatting pipeline. Visual reconstruction quality was evaluated with structural similarity index measure (SSIM), learned perceptual image patch similarity (LPIPS), and geometry-related measures. Image-space editing was applied for scene reconfiguration and evaluated by masked and non-masked SSIM to quantify local modification and global preservation. Representative scenes were integrated with operational meshes, and preliminary interaction readiness was assessed in a representative task using touch-point repeatability, point-to-plane residuals, and directional dispersion relative to the fitted interaction plane.

RESULTS:
Across 39 scenes, single-image reconstruction produced visually and spatially coherent environments, with mean input-view SSIM and LPIPS of 0.908 ± 0.005 and 0.034 ± 0.001. Geometric accuracy analysis revealed precise spatial reconstruction, with ground plane fitting error (RMSE) averaged 8.54 ± 0.24 mm. Image editing enabled geometry-preserving scene reconfiguration, with masked and non-masked SSIM of 0.582 ± 0.029 and 0.980 ± 0.003. In a representative screen-interaction task performed in mesh-integrated environments, the dominant touch cluster exhibited an approximately planar distribution aligned with the tilted screen surface, with a mean point-to-plane residual of 7.64 mm, an RMSE of 9.56 mm, and a normal-to-in-plane spread ratio of 12.2%, indicating repeatable surface contact and basic interaction readiness.

RESULTS:

Across 39 scenes, single-image reconstruction produced visually and spatially coherent environments, with mean input-view SSIM and LPIPS of 0.908 ± 0.005 and 0.034 ± 0.001. Geometric accuracy analysis revealed precise spatial reconstruction, with ground plane fitting error (RMSE) averaged 8.54 ± 0.24 mm. Image editing enabled geometry-preserving scene reconfiguration, with masked and non-masked SSIM of 0.582 ± 0.029 and 0.980 ± 0.003. In a representative screen-interaction task performed in mesh-integrated environments, the dominant touch cluster exhibited an approximately planar distribution aligned with the tilted screen surface, with a mean point-to-plane residual of 7.64 mm, an RMSE of 9.56 mm, and a normal-to-in-plane spread ratio of 12.2%, indicating repeatable surface contact and basic interaction readiness.

CONCLUSIONS:
The combination of scene reconstruction, geometry-preserving image editing, and operational mesh integration provides a practical foundation for building digital twin ophthalmic clinics for embodied AI training and validation.