Ocular Microbiology and Immunology Group
Back to OMIG Main Page

Previous | 2024 Agenda and Abstracts | Next

2024 OMIG Abstract

Evaluation of the Ocular Surface Epithelium Using Artificial Intelligence and High-Resolution Anterior Segment Optical Coherence Tomography

Jason A. Greenfield¹, Osmel Alvarez¹, Rafael Scherer¹, Diego Alba¹, Carolina Mercado¹,
Sofia De Arrigunaga¹, Douglas Rodrigues da Costa¹, William Herskowitz¹, Alessandro Jammal¹,
Michael Antonietti¹, Anat Galor^1,2, Felipe A. Medeiros¹, Carol L. Karp<sup>1

1</sup>Bascom Palmer Eye Institute, University of Miami Miller School of Medicine; ²Department of Ophthalmology, Miami Veterans Administration Medical Center, Miami, FL; ³Oregon Health & Science University, Portland, OR; ⁴Government Hospitals, Manama, Kingdom of Bahrain

Purpose: To develop a deep learning (DL) model to identify and segment ocular surface epithelium on high resolution anterior segment optical coherence tomography (AS-OCT) and discriminate ocular surface squamous neoplasia (OSSN) from pterygia/pinguecula based on epithelial thickness (ET).

Methods: A sample of 800 AS-OCT scans of 185 eyes were labeled to demarcate epithelium and used to train a DL model using FASTSAM-x. The algorithm's performance was evaluated in a separate test sample to examine the mean average precision calculated at an intersection over union above a threshold of 0.5 (mAP50). The predicted ET was then used to assign a diagnosis to biopsy-proven OSSN and non-OSSN lesions on AS-OCT scans. The area under the receiver operating characteristic curve (AUC) was assessed.

Results: The DL model had a mAP50 of 0.98. Mean predicted ET for AS-OCT scans of OSSN was 117.8μm (±49.3μm) and was significantly greater than non-OSSN at 66.8μm (±12.9μm, P<0.001). Area under the receiver operating characteristic curve for classification of OSSN was 0.902 (95%CI: 0.849-0.955) with an optimal cutoff thickness of 111.5μm.

Conclusions: A DL segmentation model applied to the ocular surface epithelium of AS-OCT scans demonstrated strong performance in differentiating OSSN from pterygia and pinguecula based on ET alone.

Disclosure:

N (JAG, OA, RS, DRC, DA, CM, SDA, WH, AJ, MA);
P (AG, PCT/US2022/029842; FM, nGoggle Inc.; CLK, PCT/US2022/029842, 63/627,578; and 63/435,503);
S (FM, Google Inc., Heidelberg Engineering, Novartis, Reichert); C (FM, AbbVie, Annexon, Astellas, Carl Zeiss Meditec, Galimedix, ONL Therapeutics, Perfusion Therapeutics, Stealth Biotherapeutics, Stuart Therapeutics, Thea Pharmaceuticals, Reichert; CLK, Interfeen Biologics, Glaukos)