Abstract

Purpose: Explore utility of Artificial Intelligence (AI) in detecting Ocular Surface Squamous Neoplasia (OSSN)
Methods: Retrospective observational study. Slit-lamp (SL) images of OSSN cases, non-OSSN surface lesions and normal ocular surfaces were collected (2013-2023). Images with minimum resolution of 1024 x 1024 pixels, under diffuse illumination were included. Data was divided into training and test sets (80:20). Deep learning (DL) algorithms were applied on images
Results: 159 images in OSSN group, 197 in non-OSSN group and 269 normal images were included. For ternary classification, MobileNet performed best, with an accuracy of 91.58%, followed by Xception and DenseNet (86.21% each). MobileNet recorded 78.5% sensitivity, 98.8% specificity, 95.6% PPV and 93.5% NPV in OSSN detection
Conclusions: AI showed good performance in image-based OSSN detection. To our best knowledge, this is the first study to establish utility of DL algorithms in OSSN detection from ocular surface images

Full Text