Memory-aware curriculum federated learning for breast cancer classification.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Diagnosis

Background:

• Mammography screening is crucial for early breast cancer detection.
• Imbalanced datasets with predominantly negative cases hinder Computer-Aided Diagnosis (CAD) system training.
• Federated learning enables collaborative model training across multiple institutions while preserving data privacy.

Purpose of the Study:

• To investigate the impact of sample presentation order on federated learning models for multi-site breast cancer classification.
• To develop a novel federated learning approach that improves the consistency and performance of local models.
• To address challenges posed by data imbalance and domain shift in federated medical image analysis.

Main Methods:

• A memory-aware curriculum learning method was developed to control the order of local model updates.
• The curriculum prioritizes training samples that are 'forgotten' after global model deployment, enhancing local model consistency.
• Unsupervised domain adaptation was integrated to handle domain shift across different clinical datasets while maintaining data privacy.

Main Results:

• The proposed memory-aware curriculum federated learning method demonstrated significant improvements in classification performance.
• Evaluation using three clinical datasets showed an average improvement of 5% in ROC-AUC and 6% in PR-AUC compared to conventional federated learning.
• An ablation study confirmed the contribution of each component of the proposed method to the overall performance enhancement.

Conclusions:

• Curriculum learning was successfully applied in a federated learning setting for the first time.
• The memory-aware curriculum federated learning approach is effective for multi-site breast cancer classification.
• The study provides a valuable contribution to improving the robustness and accuracy of AI-driven diagnostic tools in healthcare.