EyeOcuHerp Hybrid Framework: Integrating CNNs and ML for Reliable Eye (Ocular) Herpes Diagnosis

Herpes (NAGIN) on Eye / Ocular Herpes is a major cause of infectious corneal blindness, often difficult to diagnose due to overlapping morphological features. We tackle this challenge, by proposing EyeOcuHerp Hybrid Framework, blending Convolutional Neural Networks (CNNs) for feature extraction with classical Machine Learning (ML) classifiers to produce trustworthy Herpes on Eye (ocular herpes) diagnosis. We worked with 604 (279 herpes lesions and 325 without) corneal images, plus 1,400 (700 each) synthetic samples generated from CNN-based feature patterns to keep lesion categories balanced. Checking LV Prasad Eye Institute’s EMR records confirmed that both the real and synthetic features matched clinically recognized patterns: dendritic, geographic, and disciform keratitis, like branching lines seen under a slit lamp, proving their authenticity. Performance analysis revealed that, the EyeOcuHerp Model models reached AUC of 0.71 and accuracy 66%, matching classical ensembles, while keeping results interpretable through CNN based morphological features that highlight subtle texture patterns. Lesion specific stress tests showed reliably high sensitivity, above 0.86 for every morphology, while geographic and disciform types stood out with AUCs of 0.807 and 0.818. Specificity stayed moderate, 0.43 and 0.49 which shows just how hard it is to tell a herpes lesion from other corneal lesions, especially when both look equally cloudy under the slit lamp. Synthetic augmentation boosted the dataset’s variety, and a quick run of stats showed it matched real-world patterns almost perfectly. Overall, EyeOcuHerp shows practical, diagnostic process. Future studies will focus on larger datasets, sharpening specificity, and testing results to better ophthalmic care.