Author’s Disclosure Block: Weronika Jakubowska, none; Clément Playout, none; Renaud Duval, none; Evan Kalin-Hajdu, none

Abstract Body
Purpose: Periocular cancers, particularly basal cell carcinoma, are among the most common malignant eyelid lesions. As optical coherence tomography (OCT) has become increasingly integrated into ophthalmology, its potential application in oculoplastics is gaining interest. By incorporating OCT images of palpebral lesions into deep learning models, we aim to develop a universally accessible tool to assist clinicians in distinguishing between benign and malignant palpebral lesions. This study compares the diagnostic accuracy of deep learning models in classifying palpebral lesions using slit-lamp photographs and OCT imaging. Study Design: A retrospective study was conducted on OCT videos and slit-lamp photographs of palpebral lesions from patients who underwent biopsy between 2023 and 2024. The performance of deep learning classification algorithms was assessed. Methods: Two deep learning models, based on Transformer architectures adapted for 2D slit-lamp photographs and 3D OCT videos, were fine-tuned for lesion classification. Key performance metrics of the deep learning models were evaluated across a 5-fold cross-validation. The dataset consisted of both images and OCT videos, focusing on differentiating benign from malignant lesions. Results: The study analyzed 55 palpebral lesions from 50 patients, 58% of whom were female and 42% were male, with an age range of 37 to 86 years and a mean age of 69 years. Among the lesions, 60% were classified as benign and 40% as malignant. For OCT images, the model achieved an average precision of 75.8% (σ=17.3%), with a specificity of 71.9% (σ=16.1%), recall of 69.7% (σ=11.9%), an accuracy of 70.7% (σ=7.4%), and a kappa coefficient of 40.0% (σ=15%). For slit-lamp photographs, the model demonstrated stronger performance, with an average precision of 80.0% (σ=11.3%), specificity of 80.7% (σ=17.3%), recall of 76.4% (σ=20.9%), an accuracy of 77.4% (σ=11.1%) and a kappa coefficient of 54.9% (σ=20.4%). Integration of OCT and slit-lamp images into a combined model improved accuracy to 79.9% (σ=13.8%), recall to 84.9% (σ=20,4%), and kappa to 58.9% (σ=28.1%).These results show that these preliminary models demonstrate a reasonable capacity to classify palpebral lesions with promising accuracy. Conclusion: This study demonstrates the potential of deep learning models for the automated classification of palpebral lesions, especially with the novel application of OCT imaging. These models offer a non-invasive, scalable diagnostic tool that could assist in clinical evaluation and improve patient management. To our knowledge, no deep learning models currently exist that are specifically trained on OCT images of periocular lesions, positioning this work as an innovative approach to the evaluation of palpebral lesions.