Decentralized learning for medical image classification with prototypical contrastive network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Deep convolutional neural networks (CNNs) show promise in medical image classification.
• Practical application is hindered by non-independent and identically distributed (non-IID) datasets and data imbalance.
• Privacy concerns limit the use of centralized datasets from multiple institutions.

Purpose of the Study:

• To present a novel decentralized learning approach for precise medical image classification.
• To address challenges of non-IID data and data imbalance across different clients.
• To mitigate the non-IID problem using a prototypical contrastive network.

Main Methods:

• Developed a prototype contrastive network to minimize disparities among heterogeneous clients.
• Utilized an approximate global prototype to project data onto a balanced prototype space, alleviating the non-IID problem.
• Validated the algorithm using diverse datasets: EyePACS, APTOS, IDRiD (diabetic retinopathy), and COVIDx (chest X-rays).

Main Results:

• Outperformed FedAvg baseline by 3.7% in accuracy on the EyePACS dataset (balanced IID setting).
• Achieved a 6.6% accuracy enhancement over FedAvg in the Dirichlet non-IID setting on EyePACS.
• Established new state-of-the-art performance on DCC non-IID and COVID-19 datasets across multiple metrics.

Conclusions:

• The prototypical contrastive loss aligns local client data distributions with global distributions.
• An approximate global prototype addresses unbalanced data distribution by projecting data onto a balanced space.
• The model achieved state-of-the-art results on EyePACS, APTOS, IDRiD, and COVIDx datasets.