Artificial Intelligence-Based Automated Oculoplastic Measurements

Author Block: Jeremy T. Moreau , Etienne Benard-Seguin, David Plemel, Michael Ashenhurst, Ezekiel Weis, Femida Kherani, Karim Punja, Andrew Ting, Fiona Costello.  University of Calgary.

Author Disclosure Block: J.T. Moreau:  None.  E. Benard-Seguin:  None.  D. Plemel:  None.  M. Ashenhurst:  None.  E. Weis:  None.  F. Kherani:  None.  K. Punja:  None.  A. Ting:  None.  F. Costello:  None.

Abstract Title: Artificial Intelligence-Based Automated Oculoplastic Measurements

Abstract Body: Purpose:  Precise periocular measurements play an important role in presurgical planning and follow-up of patients undergoing blepharoplasty, tarsal strip canthoplasty, and other oculoplastic procedures. Classical measurement methods techniques are well-established, but can be time consuming and user dependent. Additionally, manual measurements may be impacted by patient factors such as changes in eyelid position triggered by proximity of measurement instruments or light stimuli. The purpose of our project was to design and validate a novel smartphone application for automated eyelid measurements.  Study Design:  Retrospective database, method comparison study  Methods:  We developed a novel smartphone application using artificial intelligence and image processing techniques to automatically detect eyelid margins and calculate margin to reflex distance 1 (MRD1), margin to reflex distance 2 (MRD2), palpebral fissure height/width, inter-pupillary distance, and inner/outer canthal distances. We validated the app against 120 healthy control images from the Chicago Face Database. App measurements were compared against manual photographic measurements obtained using an image analysis tool (ImageJ). Agreement was assessed using Bland-Altman difference plots. Subgroups of controls were compared to assess for equal reliability of measurements across Male, Female as well as Asian, Black, Latino, and White participants. Ethics approval was obtained from the University of Calgary Conjoint Health Research Ethics Board.  Results:  Using the current model, mean difference was 0.75mm [1.96SD: -0.39-1.89] for MRD1, -0.58 for MRD2 [1.96SD: -2.02-0.87], 0.16 [1.96SD: -1.50- 1.83] for palpebral fissure height, and -0.37 for palpebral fissure width [1.96SD: -3.00-2.26]. There were no significant differences in reliability of measurements across Male, Female or Asian, Black, Latino, and White participants.  Conclusions:  We developed a novel app for automated eyelid measurements and present data assessing the app’s reliability against manual photographic measurements. Future work will include developing and testing accuracy of additional models, and expanding app functionality to include pupil and extraocular motility measurements.