Abstract ID: 26-551
Photographic Measurement of Eyelid Height, Brow Height, and Eyelid Function: Development and Clinical Validation of a Novel Program Using a Transfer Learning–Based Deep Learning Approach
Author: BUSRA AKGUN Base Hospital / Institution: Marmara University Faculty of Medicine
Presentation Type: Oral Presentation Session: Eyelid Cosmetic & MiscellaneousDate: 11th SeptemberTime: 11.25AM
Purpose
To develop and clinically validate a computer vision system that automatically, objectively measures eyelid, brow parameters and eyelid function from photographs with millimetric precision
Methods
Standardized facial photographs of 110 subjects(healthy, various eyelid pathologies)in three gaze positions were used for model development. MediaPipe FaceMesh landmarks were manually corrected in periocular regions by an experienced ophthalmologist to build the ground truth dataset. A hybrid deep learning system refining FaceMesh predictions via delta regression was developed, using a pretrained MobileNetV3-Large backbone with anatomical heatmaps added to RGB inputs. For levator function, upgaze and downgaze images were aligned via the ECC algorithm within an ROI mask. ArUco based pixel to mm conversion enabled measurement of MRD1, MRD2, brow height, lid height, tarsal show, and levator function. For clinical validation, model measurements from 160 eyes of 80 subjects (retraction, ptosis, healthy) were compared with 2 ophthalmologists’ clinical and semi automated ImageJ measurements
Results
The system showed strong correlations with clinical and semi automated measurements for all parameters (p<0.001). Model–ImageJ correlations for brow height (ρ=0.883) and MRD2 (ρ=0.818) exceeded intra and inter observer values. Good model–ImageJ agreement was obtained for MRD1 (ICC=0.815), MRD2 (ICC=0.791), brow height (ICC=0.862), lid height (ICC=0.771) and tarsal show (ICC=0.755). For levator function, model-clinical measurements yielded ICC=0.703. In the repeatability assessment of the model, ICC values were found to be above 0.980 for all parameters
Conclusion
By computing static and dynamic eyelid function measures in a single observer independent process with millimetric precision, the system fills a critical gap in standardization and reproducibility. Transfer learning on MediaPipe FaceMesh shows general purpose landmark models can be elevated to clinical grade reliability, supporting the system as a valid alternative in routine oculoplastic assessment, preoperative planning, and follow up
Additional Authors
| First name | Last name | Base Hospital / Institution |
|---|---|---|
| Volkan | Dericioğlu | Marmara University Faculty of Medicine |