Deep Learning-Based Diagnostic Model for Ocular Surface Neoplastic Diseases - 05/05/26
Résumé |
Purpose |
To develop a deep learning (DL) model for diagnosing ocular surface tumors and evaluating its diagnostic performance.
Setting |
Development of a deep learning diagnosis algorithm.
Methods |
A total of 1491 ocular surface images representing 7 diseases—nevus (28 eyes), limbal dermoid (144), MALT lymphoma (20), ocular surface squamous neoplasia (OSSN; 138), melanoma (14), pinguecula (29), and pterygium (1,118)—were captured using slit-lamp microscopy. A YOLOv5-based DL model was trained using 5-fold cross-validation. Diagnostic performance was compared using 299 external validation images assessed by 8 corneal specialists, 7 board-certified ophthalmologists, and 8 residents.
Results |
The model achieved a positive predictive value (PPV) of 96.0%, outperforming the corneal specialists (95.0 ± 2.1%), board-certified ophthalmologists (82.6 ± 11.9%), and residents (81.9 ± 10.7%). Disease-specific PPVs were: nevus 75.0%, limbal dermoid 93.5%, MALT lymphoma 57.9%, OSSN 87.0%, melanoma 38.5%, pinguecula 82.1%, and pterygium 96.7%. The area under the curve (AUC) was: nevus 0.897 (95% confidence interval [CI], 0.810-0.983), limbal dermoid 0.998 (95% CI, 0.996-1.000), MALT lymphoma 0.894 (95% CI, 0.794-0.993), OSSN 0.954 (95% CI, 0.933-0.975), melanoma 0.966 (95% CI, 0.919-1.000), pinguecula 0.954 (95% CI, 0.912-0.995), and pterygium 0.984 (95% CI, 0.976-0.992). Sensitivities were: nevus 0.643, limbal dermoid 0.993, MALT lymphoma 0.550, OSSN 0.681, melanoma 0.357, pinguecula 0.793, and pterygium 0.991. Specificities were: nevus 0.996, limbal dermoid 0.993, MALT lymphoma 0.995, OSSN 0.990, melanoma 0.995, pinguecula 0.997, and pterygium 0.898.
Conclusions |
The deep learning model demonstrated high diagnostic accuracy for common ocular surface tumors such as pterygium and limbal dermoid, while diagnostic performance for rare malignancies, including melanoma and MALT lymphoma, remains limited and requires further refinement.
Synopsis |
We developed a deep learning model that demonstrated promising performance in identifying ocular surface neoplastic diseases, suggesting its potential as a supportive diagnostic tool in ophthalmic practice.
Le texte complet de cet article est disponible en PDF.Plan
Vol 286
P. 184-195 - juin 2026 Retour au numéro
Article précédent
- Systemic Immunosuppressant Cessation/Tapering Linked to MGD Progression Post-HSCT
- Wenxin Zhao, Xiaohui Luo, Jing Yang, Yinglin Liao, Ling Jin, Ziyan Chen, Lingyi Liang
- Adaptive Temporal Mixture of Experts for Predicting Stiffness Metrics From the Ocular Response Analyzer and Identifying Keratoconus
- Yue Zhang, Swati Padhee, Phillip T. Yuhas, Cynthia J. Roberts, Srinivasan Parthasarathy
Bienvenue sur EM-consulte, la référence des professionnels de santé.
L’accès au texte intégral de cet article nécessite un abonnement.
Déjà abonné à cette revue ?