Extraction of thin orthokeratology lens surfaces in spectral-domain OCT using prior-guided signed-distance regression

Accurate extraction of inner surfaces of dry, isolated orthokeratology (OK) lenses from spectral-domain optical coherence tomography (SD-OCT) images is challenging because the target is a thin reflective interface affected by speckle noise, signal attenuation, and reflection artifacts. Missing segments can leave gaps across radial views, making surface coverage an important consideration alongside localization accuracy. We propose PG-SDFNet, which combines continuous signed-distance field (SDF) regression with a coarse geometric prior generated automatically from each OCT B-scan. The prior is incorporated through input-level fusion and prior-guided skip modulation to guide surface representation learning, while annotated contours provide dense distance supervision. A cohort of 839 acquisitions across seven OK-lens brands yielded 53,696 radial B-scans. A single-brand subset of 95 retained acquisition UUIDs, comprising 5,737 annotated B-scans, was used for model development and internal evaluation. On the test set of 19 acquisition UUIDs (1,140 B-scans), PG-SDFNet achieved mean acquisition-level recall of 0.9405 and Dice of 0.4346. Relative to SDF-Concat, recall was higher in all 19 acquisitions and Dice in 16. Radial completeness reached 0.9029 at a per-view recall threshold of 0.80. In this single-brand internal evaluation, PG-SDFNet recovered extensive surface profiles across the acquired radial views, providing potential inputs to subsequent geometric analysis.