Automated diagnosis of keratitis from low-quality slit-lamp images using an improved generative adversarial network

Keratitis is a major cause of visual impairment worldwide, with timely and accurate diagnosis being critical to preventing irreversible vision loss. Existing deep learning (DL) models rely heavily on high-quality slit-lamp (HQS) images, limiting their applicability in real-world clinical scenarios where low-quality slit-lamp (LQS) images are frequently encountered due to acquisition-related artifacts. Here, we propose an automated keratitis diagnosis framework for LQS images (AKDF-LQSI) that integrates image enhancement with lesion-aware classification. A random degradation algorithm is designed to simulate common clinical artifacts, enabling paired LQS-HQS training. The framework incorporates an improved generative adversarial network to restore diagnostic features and a context-aware classifier to enhance lesion representation. AKDF-LQSI demonstrates robust performance when evaluated on a multicenter dataset of 10,498 slit-lamp images, achieving an AUC exceeding 0.949. This approach presents a practical solution for keratitis diagnosis from degraded images, with strong potential to reduce diagnostic disparities in resource-limited environments.