Automated Detection Of Keratorefractive Laser Surgeries On Optical Coherence Tomography Using Deep Learning
Published 2024 - 42nd Congress of the ESCRS
Reference: PO271 | Type: Poster | DOI: 10.82333/hkew-8z67
Authors: Jad F. Assaf* 1 , Hady Yazbeck 2 , Dan Reinstein 3 , Timothy Archer 3 , Roland Assaf 2 , Diego de Ortueta 4 , Juan Arbelaez 5 , Maria Clara Arbelaez 5 , Shady Awwad 6
1Casey Eye Institute,Oregon Health & Science University,Portland,United States;Faculty of Medicine,American University of Beirut,Beirut,Lebanon, 2Faculty of Medicine,American University of Beirut,Beirut,Lebanon, 3Reistein Vision,Londond,United Kingdom, 4Aurelios Augenlaserzentrum Recklinghausen,Recklinghausen,Germany, 5Muscat Eye Laser Center,Muscat,Oman, 6Department of Ophthalmology,American University of Beirut,Beirut,Lebanon
Purpose
To report a deep learning neural network on anterior segment optical coherence tomography (AS-OCT) for automated detection of different keratorefractive laser surgeries—including Laser In-Situ Keratomileusis with femtosecond microkeratome (Femto-LASIK), LASIK with mechanical microkeratome, photorefractive keratectomy (PRK), small incision lenticule extraction (SMILE), and non-operated eyes—while also distinguishing the targeted ametropias, such as myopic and hyperopic treatments, within these procedures..
Setting
This study was conducted in the ophthalmology department at the American University of Beirut, Beirut, Lebanon.
Methods
A total of 14,948 eye scans from 2,278 eyes of 1,166 subjects were used to develop a deep learning neural network algorithm with an 80/10/10 patient distribution for training, validation, and testing phases, respectively. The algorithm was evaluated for its accuracy, F1-scores, area under precision-recall curve (AUPRC), and area under receiver operating characteristic curve (AUROC).
Results
On the test dataset, the neural network was able to detect the different surgical classes with an accuracy of 96%, a weighted-average F1-score of 96% and a macro-average F1-score of 96%. The neural network was further able to detect hyperopic and myopic subclasses within each surgical class, with an accuracy of 90%, weighted-average F1 score of 90%, and macro-average F1-score of 83%.
Conclusions
Determining a patient's keratorefractive laser history is vital for customizing treatments, performing precise intraocular lens (IOL) calculations, and enhancing ectasia risk assessments, especially when electronic health records are incomplete or unavailable. Neural networks can be used to accurately classify keratorefractive laser history from AS-OCT scans, a step in transforming the AS-OCT from a diagnostic to a screening tool in the refractive clinic.