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

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Infectious keratitis (IK) is a major cause of blindness globally.
• Distinguishing IK subtypes is clinically challenging.
• Current deep learning (DL) methods often use single imaging modalities, missing crucial diagnostic information.

Purpose of the Study:

• To develop a novel DL framework, KeraFusionNet, for accurate IK diagnosis and subtype differentiation.
• To integrate multiple slit-lamp imaging (SLI) modalities for enhanced diagnostic capabilities.
• To overcome the limitations of single-modality DL approaches in IK diagnosis.

Main Methods:

• KeraFusionNet processes three imaging modalities: diffuse white light, slit beam, and cobalt blue light with fluorescein staining.
• The framework employs modality-specific expert subnetworks.
• A dynamic gating network adaptively fuses feature representations for classification.
• Validation was performed on multicenter datasets from Zhejiang Eye Hospital (ZEH) and Aier Guangming Eye Hospital (AGEH).

Main Results:

• KeraFusionNet achieved 87.40% overall accuracy on the ZEH dataset.
• AUCs reached 0.9899 for healthy cornea, 0.9177 for herpes simplex keratitis (HSK), 0.9653 for bacterial keratitis (BK), 0.9803 for fungal keratitis (FK), and 0.9635 for other abnormalities.
• The model demonstrated 83.49% accuracy on the independent AGEH dataset.

Conclusions:

• KeraFusionNet effectively integrates complementary diagnostic information from routine SLI modalities.
• The multimodal fusion framework shows effectiveness and generalizability for automated IK diagnosis.
• This approach offers a promising tool for precise clinical decision-making in ophthalmic practice.