More Than Meets the Eye: Multi-Modal Deep Learning for Classifying Eyelid Lesions
Author: Weronika Jakubowska
Base Hospital / Institution: Université de Montréal
ePoster presentation
Abstract ID: 25-382
Purpose
While OCT is increasingly used in ophthalmology, its role in eyelid lesion assessment is still emerging. This study aims to develop a universally accessible, non-invasive deep learning tool to assist clinicians in distinguishing benign from malignant palpebral lesions. We evaluate and compare the diagnostic performance of models trained on slit-lamp photographs, OCT imaging, and their combination.
Methods
In this retrospective study, 56 palpebral lesions from 50 patients (mean age: 69 years; 58% female) who underwent biopsy between 2023 and 2025 were included. Lesions were imaged using slit-lamp photography and swept-source OCT. Three deep learning models were developed: a Vision Transformer (ViT) for 2D slit-lamp photographs, a Video Masked Autoencoder (VideoMAE) for 3D OCT volumes, and a novel fusion model combining both modalities. 5-fold cross-validation was used with extensive regularization.
Results
Of the 56 palpebral lesions analyzed, 57% were benign and 43% malignant. The OCT model achieved 84% average precision (σ=7.8%), 78.6% specificity (σ=5.8%), 68% recall (σ=20.4%), 73.2% accuracy (σ=5.8%), and a kappa of 45.4% (σ=13%). The slit-lamp model showed 87.7% precision (σ=16%), 81% specificity (σ=18.9%), 84% recall (σ=19.6%), 82.1% accuracy (σ=5.8%), and kappa of 64% (σ=10.9%). The combined model performed best with 91.2% precision (σ=6.4%), 83.3% accuracy (σ=5.3%), 81% specificity (σ=18.9%), 86% recall (σ=12.7%), and kappa of 66.7% (σ=10.2%).
Conclusion
This study validates the potential of deep learning for non-invasive classification of palpebral lesions, with the integration of OCT and slit-lamp data significantly enhancing diagnostic performance. To our knowledge, this is the first model trained specifically on OCT volumes of periocular tumors, marking a novel contribution to oculoplastic diagnostics. As a proof of concept, it underscores the potential utility of deep learning models applied to standard, widely available OCT systems in clinical practice. Such tools could improve access to care in remote areas, facilitate triage, and expedite treatment of malignant lesions.
Additional Authors
| First name | Last name | Base Hospital / Institution |
|---|---|---|
| Clément | Playout | Université de Montréal |
| Evan | Kalin-Hajdu | Université de Montréal |
