Dual U-Net with multi-task attention for automated eyelid curvature quantification

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Eyelid curvature is crucial for diagnosing ophthalmic conditions and evaluating surgical outcomes.
• Accurate and reproducible eyelid margin analysis is essential for clinical practice.

Purpose of the Study:

• To develop an automated image processing method for eyelid margin curve extraction.
• To perform quantitative curvature analysis for ophthalmic applications.

Main Methods:

• A dual-branch U-Net architecture (AtDU-Net) was employed for simultaneous segmentation of palpebral fissure and corneal regions.
• Eyelid margin curves were extracted from segmented images and fitted using second-order polynomials.
• Quantitative curvature values were calculated based on the fitted curves.

Main Results:

• The AtDU-Net model achieved high segmentation performance with IoU of 0.979 and Dice coefficient of 0.989.
• Automated eyelid curvature measurements demonstrated strong correlation with manual annotations (r=0.9032 upper, r=0.9154 lower).
• Bland-Altman analysis confirmed high consistency, with over 92% of samples within agreement limits.

Conclusions:

• The proposed automated method offers superior accuracy, robustness, and consistency compared to manual measurements.
• This technique provides reliable support for eyelid morphological analysis and surgical planning in clinical settings.