Privacy-aware diabetic retinopathy grading and visual lesion-focused interpretability through mixture-of-experts...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Diabetic Retinopathy (DR) is a leading preventable cause of vision loss.
• Automated screening systems are needed to analyze medical data across institutions without centralizing sensitive information.
• Federated Learning (FL) enables collaborative model training while maintaining data privacy, but faces challenges with non-IID data, communication overhead, and convergence stability in medical imaging.

Purpose of the Study:

• To propose and evaluate a federated Mixture-of-Experts (FL-MoE) framework for Diabetic Retinopathy classification.
• To combine interpretable deep learning with expert specialization to address FL limitations in medical imaging.
• To assess the performance of various backbone architectures (CNN, CNN-LSTM, ViT) within the FL-MoE framework using public DR datasets.

Main Methods:

• Developed a federated Mixture-of-Experts (FL-MoE) framework for DR classification.
• Evaluated multiple deep learning backbone architectures (CNN, CNN-LSTM, Vision Transformer) within the FL-MoE setup.
• Utilized the EyePACS and APTOS-2019 retinal fundus datasets for training and validation.
• Employed Grad-CAM for explainability analysis and Intersection-over-Union (IoU) for localization quality assessment.

Main Results:

• The FL-MoE framework demonstrated improved performance under heterogeneous data distributions across different backbone architectures.
• The CNN-LSTM backbone achieved 76.2% accuracy and 89.5% AUC on the EyePACS dataset, with significantly reduced communication costs compared to transformer models.
• CNN-LSTM exhibited more stable convergence and greater robustness to client-level data heterogeneity.
• Explainability analysis showed attention maps highlighting relevant retinal regions, though localization accuracy (mean IoU < 0.03) was coarse.

Conclusions:

• The proposed FL-MoE framework with a CNN-LSTM backbone provides an effective and scalable solution for privacy-aware Diabetic Retinopathy screening.
• This approach outperforms standard federated baselines and personalized FL methods under heterogeneous data conditions.
• The framework offers a practical method for federated clinical environments, enhancing diagnostic capabilities while preserving patient data privacy.