Adaptive Multimodal Fusion in Vertical Federated Learning for Decentralized Glaucoma Screening

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning
• Data Privacy

Background:

• Early glaucoma detection is crucial for preventing vision loss.
• Unimodal retinal imaging has limitations in accuracy and context.
• Fragmented clinical data and privacy regulations hinder centralized AI development.

Purpose of the Study:

• To develop a decentralized multimodal glaucoma detection framework.
• To address data sparsity and privacy challenges in glaucoma diagnosis.
• To improve the accuracy and robustness of glaucoma screening.

Main Methods:

• Proposed a Quality Aware Vertical Federated Learning (QAVFL) framework.
• Integrated clinical text, retinal images, and biomedical signals using modality-specific encoders.
• Employed a Fusion Attention Module (FAM) for adaptive multimodal weighting.
• Utilized homomorphic encryption and differential privacy for secure aggregation.

Main Results:

• QAVFL achieved 98.6% accuracy, 98.6% recall, 97.0% F1-score, and 0.992 AUC.
• Demonstrated statistically significant improvements over unimodal and early fusion methods (p < 0.01).
• Validated performance in heterogeneous non-IID settings.

Conclusions:

• Dynamic multimodal fusion is effective for privacy-preserving decentralized learning.
• QAVFL offers a scalable and clinically applicable solution for glaucoma screening.
• The framework addresses challenges of fragmented healthcare data for AI-driven diagnostics.