Explainable Federated Medical Image Analysis Through Causal Learning and Blockchain.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated learning (FL) facilitates collaborative model training on distributed private medical data.
• Challenges in FL for medical image analysis include high communication overhead and data heterogeneity.
• Existing FL methods lack sufficient interpretability and trust for clinical deployment.

Purpose of the Study:

• To propose novel FL techniques integrated with explainable artificial intelligence (XAI) for efficient, accurate, and trustworthy medical image analysis.
• To address communication overhead and data heterogeneity challenges in FL for medical imaging.
• To enhance the clinical viability of FL through improved interpretability and trust.

Main Methods:

• Implemented a heterogeneity-aware causal learning approach to selectively sparsify model weights based on causal contributions, reducing communication.
• Utilized blockchain for decentralized quality assessment of client datasets, informing aggregation weights.
• Integrated XAI to provide essential model explanations and trust mechanisms.

Main Results:

• The XAI-integrated FL framework demonstrated enhanced efficiency, accuracy, and interpretability.
• Causal learning significantly decreased communication overhead while maintaining medical image segmentation accuracy.
• Blockchain-based data valuation effectively mitigated performance degradation from low-quality local datasets.

Conclusions:

• The proposed XAI-integrated FL framework offers an efficient, accurate, and trustworthy solution for medical image analysis.
• Causal learning and blockchain-based data valuation are effective strategies for overcoming FL challenges in healthcare.
• The framework's interpretability and trust mechanisms pave the way for broader clinical adoption of federated learning in medical imaging.