Explainable Artificial Intelligence–Assisted Differentiation Of Pterygium And Conjunctival Intraepithelial Neoplasia Using Slit-Lamp Photographs: A Pilot Diagnostic Study
Published 2026 - 44th Congress of the ESCRS
Reference: POCOR462 | Type: Free paper | DOI: 10.82333/ymsk-1m71
Authors: Dr Yurdagul Girgin 1 , Dr Gozde Sahin Vural* 2
1Corlu State Hospital,Tekirdag,Türkiye, 2Balikesir University,Balikesir,Türkiye
Purpose
To evaluate the feasibility of an explainable artificial intelligence (AI)–assisted framework for differentiating pterygium from conjunctival intraepithelial neoplasia (CIN) using slit-lamp anterior segment photographs.
Setting
This pilot study was conducted at a single tertiary ophthalmology center using slit-lamp photographs obtained during routine clinical examinations. Images of ocular surface lesions diagnosed as pterygium or CIN were retrospectively collected and analyzed within a standardized evaluation framework.
Methods
In this pilot diagnostic study, slit-lamp photographs of biopsy-proven pterygium and conjunctival intraepithelial neoplasia (CIN) were retrospectively collected from a tertiary ophthalmology clinic and recorded in a standardized evaluation table. An explainable AI-assisted protocol based on predefined morphological criterias and suspicious OSSN findings were applied. Images were divided into calibration and blinded test datasets. AI diagnoses were compared with reference standard diagnoses based on histopathological examination. Data organization and structured annotation were supported using ChatGPT (OpenAI, GPT-5.2) as an assistive analytical tool, while all clinical decisions were made by the researchers.
Results
A total of 40 slit-lamp images were analyzed. According to AI primary classification, 20 lesions were categorized as pterygium and 20 as CIN. Pterygium cases were predominantly characterized by triangular fibrovascular proliferation with organized radial vascular patterns and smooth translucent surfaces. In contrast, CIN lesions demonstrated irregular or gelatinous morphology, leukoplakia, nodularity, and prominent feeder vessels. In cases where reference standard diagnosis was available, AI predictions showed full agreement with histopathological diagnosis within this pilot dataset. The explainable feature-based framework enabled transparent interpretation of AI decision patterns rather than black-box classification.
Conclusions
This pilot study demonstrates the feasibility of an explainable AI-assisted approach for differentiating pterygium from CIN using slit-lamp photographs. Morphology-based AI interpretation may assist clinicians in identifying ocular surface neoplasia during routine examinations. Larger studies with expanded datasets and histopathological confirmation are required to determine diagnostic accuracy and clinical applicability.