Developing a privacy-preserving deep learning model for glaucoma detection: a multicentre study with federated...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning (DL) shows promise for glaucoma detection.
• Patient privacy and data security are significant challenges in pooling data for DL models.
• A privacy-preserving DL model using federated learning (FL) was developed for glaucoma detection from optical coherence tomography (OCT) images.

Purpose of the Study:

• To develop and evaluate a privacy-preserving federated learning (FL) model for glaucoma detection using OCT images.
• To address patient privacy and data security concerns in multi-center AI model development.
• To assess the generalizability and stability of the FL model across different eye centers and datasets.

Main Methods:

• A multi-center study involving seven eye centers across Hong Kong, the USA, and Singapore.
• Utilized the federated learning (FL) paradigm with a central server and local model training at each center.
• Employed the FedProx algorithm for model parameter aggregation and redistribution, experimenting with three 3D networks and testing on unseen datasets.

Main Results:

• The FL model demonstrated consistent performance across seven centers, with accuracies ranging from 78.3% to 98.5%.
• Stable performance was observed on two unseen datasets, with accuracies between 81.3% and 87.8%.
• The FL model achieved non-inferior performance compared to traditional models and outperformed individual center models.

Conclusions:

• The 3D FL model effectively leveraged multi-center datasets without data exchange, ensuring patient privacy and data security.
• This OCT-based FL paradigm demonstrates a viable pathway for real-world application of AI in glaucoma detection.
• The study highlights the potential of federated learning for advancing artificial intelligence in ophthalmology.