A Mixture-of-Experts Network for Infectious Keratitis Classification Using Multimodal Slit-Lamp Images: A Multicenter Study.

Purpose: Infectious keratitis (IK) remains one of the leading causes of corneal blindness worldwide, and subtype distinguishing continues to pose significant clinical challenges. Existing deep learning (DL) approaches typically rely on a single imaging modality, which overlooks the distinct diagnostic cues present in other modalities. Methods: To address this limitation, we propose KeraFusionNet, a novel DL framework that simultaneously processes three imaging modalities: diffuse white light, slit beam, and cobalt blue light with fluorescein staining. The model integrates three modality-specific expert subnetworks and uses a dynamic gating network to adaptively fuse their feature representations for final classification. We validated the method using a multicenter dataset comprising 3236 images from 820 patients at Zhejiang Eye Hospital (ZEH) and 862 images from 261 patients at Aier Guangming Eye Hospital (AGEH). Results: On the ZEH dataset, the model achieved an overall classification accuracy of 87.40%, with areas under the curve (AUCs) of 0.9899, 0.9177, 0.9653, 0.9803, and 0.9635 for a healthy cornea, herpes simplex keratitis (HSK), bacterial keratitis (BK), fungal keratitis (FK), and other cornea abnormalities, respectively. On the independent AGEH dataset, the accuracy reached 83.49%. Conclusions: These results demonstrate the effectiveness and generalizability of our multimodal fusion framework, offering a promising tool for automated IK diagnosis and subtype differentiation. Translational Relevance: By integrating complementary diagnostic information from routine slit-lamp imaging (SLI) modalities, KeraFusionNet enables accurate, automated differentiation of IK subtypes, supporting more precise clinical decision making in real-world ophthalmic practice.