Physics-guided neural reconstruction of corneal topography from Placido-ring images

Placido-ring corneal topography is essential for refractive surgery planning, keratoconus screening, and postoperative monitoring, but reconstructing corneal shape is difficult because it is a nonlinear specular-reflection inverse problem and is vulnerable to noise, tear-film breakup, and missing reflections. We propose physics-guided neural reconstruction (PGNR), a convolutional framework that reconstructs corneal elevation from polar Placido measurements while enforcing meridional geometric-optics forward-model consistency through a measurement-space loss that reduces nonlinear noise bias. Under controlled degradation, PGNR reduces the elevation mean absolute error (MAE) by 69% versus forward ray-tracing, while its elevation MAE increases by only 5.1% with 20% missing data.