Highly efficient homomorphic encryption-based federated learning for diabetic retinopathy classification.

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness in working-age adults globally.
• Machine learning (ML) shows promise for DR diagnosis, but requires diverse data for generalizability.
• Federated learning (FL) enables decentralized training but faces privacy concerns like gradient inversion attacks.

Purpose of the Study:

• To develop and test a computationally efficient FL framework integrating homomorphic encryption (HE) for privacy preservation in DR screening.
• To safeguard patient data against privacy breaches during decentralized ML model training.

Main Methods:

• Utilized a dataset of 6457 retinal fundus images from APTOS-2019 and ODIR-5K.
• Employed RETFound, a pre-trained foundation model, for feature extraction from distributed retinal images.
• Trained a lightweight multiclass logistic regression head (MLRH) model using FL with encrypted features and HE for privacy.

Main Results:

• The FL-trained MLRH model achieved performance comparable to a fully fine-tuned model on centralized data (AUCs of 0.93 and 0.78).
• Demonstrated significant efficiency gains: 95.9-fold reduction in computation time and 63.0-fold reduction in data transfer.
• Confirmed that integrated HE effectively protected patient data against gradient inversion attacks.

Conclusions:

• Advanced privacy-preserving ML-based DR screening technology.
• The framework supports equitable vision care by enabling robust and secure DR diagnosis.
• Highlights the potential of combining FL and HE for secure medical data analysis.