Deep Learning-Based Detection of Ocular Surface Squamous Neoplasia from Ocular Surface Images.

Abstract

Introduction: Ocular surface squamous neoplasia (OSSN) is a broad entity encompassing a spectrum of squamous neoplasms of conjunctiva and cornea. This study aimed to explore the utility of artificial intelligence (AI) models in detecting OSSN from slit-lamp (SL) images.

Methods: This is a retrospective observational study. SL images of OSSN disease, non-OSSN ocular surface lesions (OOSD), and normal ocular surfaces (N) were collected (2013-2023). Images with minimum resolution of 1,024 × 1,024 pixels under diffuse illumination were included. Data were divided into training and testing sets (85:15). Deep learning (DL) algorithms were applied for ternary classification of the SL images (OSSN, OOSD, and normal). Three AI models - MobileNetV2, Xception, and DenseNet121 - were used in the study. A fivefold cross-validation strategy was utilized for robust model evaluation.

Results: A total of 163 images in OSSN group, 202 in OOSD group, and 269 normal ocular surface images were included (n = 634). Data augmentation was performed to increase and balance the data. The average accuracies for OSSN detection for DenseNet121, MobileNetV2, and Xception were 83%, 88.8%, and 84.5%, respectively. MobileNetV2 and Xception had a similar average sensitivity for OSSN detection (74%) while MobileNetV2 was the most specific DL algorithm (96.25%) for OSSN detection.

Conclusions: AI models showed good performance in image-based OSSN detection. AI models may provide a promising tool for OSSN screening in primary health care centers and for teleconsultation from remote areas in the future.