A comprehensive system for eyelid analysis using deep learning: automated measurement of eyelid position and corneal exposure

Abstract Quantitative eyelid analysis is crucial for diagnosing and managing eyelid disorders, such as ptosis, and for surgical planning. However, traditional manual measurements of Marginal Reflex Distance (MRD) and Corneal Exposure Ratio are often error-prone and inconsistent. Existing deep learning approaches using RGB imaging face challenges in accurately delineating eye structures due to insufficient contrast and varying iris pigmentation. This study introduces a comprehensive system integrating infrared (IR) imaging with deep learning-based eye segmentation to overcome these limitations. The system utilizes an IR camera with facial rotation compensation for consistent image acquisition and employs deep learning algorithms optimized for IR imaging characteristics. A dataset of 2,121 IR images from 303 patients (internal validation) and 406 public dataset images (external validation) was used. Multiple models including RITnet and SegFormer were evaluated, with data augmentation techniques applied to improve performance. SegFormer with data augmentation demonstrated the highest performance, achieving mean IoU of 0.9325 for internal validation and 0.9288 for external validation. Our automated system achieved improved precision compared to manual measurements, with MRD1 MSE of 0.362±0.511 mm and MRD2 MSE of 0.286±0.324 mm. This proposed system offers a robust, efficient, and standardized approach for quantitative eyelid analysis, leveraging IR imaging advantages to address the shortcomings of conventional assessment methods and providing clinical-grade accuracy for automated eyelid parameter measurement.