Deep-Learning Model for Iris and Eyebrow Segmentation and Automation of Eye Landmark Measurements in Acquired Ptosis

Background: Acquired ptosis is a common eyelid disorder in elderly patients, causing visual disturbance and cosmetic concerns. Accurate evaluation of periocular anatomy, including eyebrow and iris position, is essential for surgical planning, but current manual assessments are time-consuming and subjective. Objectives: This study aimed to develop deep-learning models for iris and eyebrow segmentation to automate eye landmark measurements and enable objective, standardized analysis in patients with acquired ptosis. Methods: We retrospectively collected 612 facial images from 209 ptosis patients. Images were labeled for iris and eyebrow segmentation and split into training, validation, and test sets (8:1:1). A deep-learning model was developed to automatically segment the iris and eyebrow regions and automatically measure seven landmarks: MRD1, MRD2, medial eyebrow end, medial limbus, pupil center, lateral limbus, and lateral eyebrow end. Results: The iris segmentation model achieved accuracy of 99.7%, precision of 97.6%, recall of 98.3%, an F1 score of 97.9%, and intersection over union of 95.9%. The corresponding metrics for the eyebrow segmentation model were 98.6%, 92.6%, 91.5%, 91.5%, and 85.0%. The mean absolute percentage error and root mean square error for the automated landmark measurements were 4.00% and 2.48 mm, respectively. Conclusions: The high performance of the segmentation models and the automated measurements supports their potential use for objective and standardized analyses of acquired ptosis. These findings may aid the future development of predictive tools for use in surgical planning.