Detection of Ocular Surface Squamous Neoplasia Using Artificial Intelligence with High-Resolution Anterior Segment Optical Coherence Tomography

Abstract

Purpose : To develop and validate a deep learning (DL) model to differentiate ocular surface squamous neoplasia (OSSN) from clinically confounding conditions, such as pterygium and pinguecula, using high resolution anterior segment optical coherence tomography (HR-OCT) images.

Methods : Imaging data was extracted from Optovue HR-OCT (Fremont, CA) and subjects’ clinical or biopsy-proven diagnoses were collected from electronic medical records. A DL classification model was developed using two methodologies: (1) an autoencoder model was trained with unlabeled data from 84,687 HR-OCT images of 5892 eyes to reduce the image into a low-dimensional latent space, and (2) a Vision Transformer supervised model (ViT) used labeled data as an input for fine-tuning a binary classifier (OSSN vs. non-OSSN lesions). A sample of 435 eyes from 308 subjects (1,223 HR-OCT) images were classified by expert graders into “OSSN or suspicious for OSSN” and “pterygium or pinguecula”. The best performing algorithm was selected using the accuracy of the validation dataset. The algorithm's diagnostic performance was then assessed in a separate test sample using 679 scans (85 eyes, 48 subjects) with biopsy-proven OSSN. Analysis was conducted at the patient level and, for the DL model, an eye was classified as having OSSN if at least one scan was positive for OSSN.

Results : The model had an accuracy of 91.7% (95%CI: 80.0-97.7%), with sensitivity of 90.9% (95%CI: 70.8-98.9%) and specificity of 92.3% (95%CI: 74.9-99.1%). The area under the receiver operating characteristic curve (AUC) was 0.916 (95%CI: 0.835-0.997) for the best performing DL model at the patient level. Human gradings had lower performance than the DL model (AUC=0.687, P=0.003) at the patient level, with sensitivity of 68.2% (95%CI: 45.1-86.1) and specificity of 69.2% (95% CI: 48.2-85.7%).

Conclusions : A DL model, applied to HR-OCT scans of the anterior segment, demonstrated high accuracy in differentiating OSSN from similar clinical conditions such as pterygium and pinguecula. Interestingly, the model surpassed the diagnostic performance of clinicians in this study and shows promise for enhancing clinical decision-making. Further research is warranted to explore the integration of this AI-driven approach in routine screening and diagnostic protocols for OSSN.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.