VisionEyeNet: a customized deep learning framework for early diagnosis of keratitis and uveitis

Abstract

Keratitis and uveitis are increasingly prevalent ocular disorders, often linked to delayed detection and limited specialist access, particularly in rural healthcare settings. These diseases can lead to severe visual impairment or irreversible blindness if not identified at an early stage. Traditional diagnostic approaches are manual, time-consuming, and prone to human error, making them challenging for large-scale screening. To address these limitations, this study presents VisionEyeNet, a framework for automatic classification of keratitis and uveitis. VisionEyeNet integrates MobileNetV2 and DenseNet121 within a fusion architecture, along with image enhancement methods such as adaptive gamma correction and specular reflection suppression. The model was trained and evaluated on a curated dataset of 1,860 slit-lamp images (960 uveitis and 900 keratitis) using a patient-wise split (71.5% training, 8.4% validation, and 20% testing). On the independent test set, it achieved 98.0% accuracy (95% CI: 97.1–98.8%) with balanced performance across classes. Inference analysis showed an average processing time of 51±2 ms per image, supporting real-time use. These results indicate that VisionEyeNet has strong potential as a clinically useful decision-support tool, particularly in resource-limited settings.