Multimodal Distillation and Fusion for Enhanced Age-Related Macular Degeneration Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Age-related macular degeneration (AMD) is a primary cause of vision loss in older adults.
• Accurate AMD diagnosis and subtyping are crucial for effective clinical management.
• Current methods struggle to integrate multimodal imaging data due to complexity and data variability.

Purpose of the Study:

• To develop a novel Multimodality Distillation and Fusion Network (ModiF) for robust AMD classification.
• To enhance the utilization of complementary information from diverse imaging modalities.
• To improve diagnostic accuracy and subtyping of AMD.

Main Methods:

• Designed modality-specific feature extractors for OCT, OCTA, and CFP images.
• Implemented cross-modal soft supervision for knowledge distillation.
• Utilized a principal-modality guided distillation for cross-modality transfer.
• Introduced a confidence-aware fusion module for adaptive modality weighting.

Main Results:

• ModiF demonstrated superior performance compared to state-of-the-art methods.
• Consistent outperformance across internal (OCT, OCTA) and public (CFP, OCT) datasets.
• Validated the effectiveness and generalization capability of the proposed network.

Conclusions:

• The ModiF network offers a robust solution for AMD classification by effectively integrating multimodal imaging data.
• The proposed distillation and fusion strategies enhance feature discriminability and diagnostic accuracy.
• ModiF shows significant potential for improving clinical decision-making in AMD management.