Intelligent interferometric analysis of lipid layer thickness for clinical evaluation of dry eye disease

Accurate and reproducible measurement of tear film lipid layer thickness (LLT) is essential for evaluating meibomian gland dysfunction (MGD) and diagnosing dry eye disease (DED). However, existing interferometric systems mainly rely on qualitative or semi-quantitative assessments, which limit their clinical utility. This study aimed to develop and validate a fully automated and deep learning framework for quantitative LLT analysis from white-light interference images.

Preclinical images from BALB/c mice and clinical datasets from healthy volunteers and patients with MGD were analyzed. Eight deep learning segmentation models, including U-Net, DeepLabV3+, and Unet++, were compared to identify the optimal architecture. Following segmentation, a physics-informed optical mapping approach was implemented to convert pixel intensity values into quantitative LLT measurements. Model performance and LLT accuracy were validated against manually annotated ground truth data.

Unet++ achieved the highest segmentation performance, with Dice scores of 0.98 for animal and 0.99 for human datasets. Quantitative LLT values derived from Unet++ showed strong linear correlation with ground truth measurements (R² = 0.972 in animals, R² = 0.994 in humans). The system consistently produced quantitative LLT measurements that distinguished between normal and pathological tear films, demonstrating robustness across datasets and imaging conditions.

This study presents the fully automated LLT quantification framework that integrates interferometry, deep learning, and physics-based modeling. By eliminating manual grading, the proposed system offers a robust, scalable, and clinically applicable tool for objective tear film evaluation, supporting both routine diagnosis and longitudinal monitoring of DED and MGD.