Optimization of sensitivity, resolution, and phase-retardance stability in all-fiber Jones matrix optical coherence tomography system

Keratoconus (KC) is a progressive corneal condition causing thinning and bulging, leading to visual distortion. Early detection is crucial to prevent complications like scarring or the need for corneal transplants. This study leverages a custom ultrahigh-resolution Polarization-Sensitive Optical Coherence Tomography (UHR-PS-OCT) system to develop an AI model for classifying healthy, subclinical keratoconus (SKC), and KC eyes, aiming to enable early detection. Using data from 255 eyes (101 healthy, 105 KC, 49 SKC), a random forest model achieved an AUROC of 0.911, with accuracy, recall, and precision of 0.820, 0.820, and 0.804, respectively. While KC classification was highly accurate (98.1%), SKC detection showed limitations. Novel features, such as phase retardation (PR) and spatial mapping of epithelial and Bowman’s layer thickness, enabled early KC detection without corneal topography. The model autonomously identified critical spatial patterns, including localized thinning and PR distortions. However, overlaps in ultra-structural features between healthy and suspicious corneas reduced SKC accuracy. Future studies will refine and validate this approach with larger cohorts.