Federated Learning Framework for Privacy-Preserving Explainable AI-Driven Clinical Decision-Making.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Centralized AI in diagnostics faces challenges with patient data privacy, heterogeneity, and generalizability.
• Existing frameworks struggle with secure and adaptable clinical decision support in distributed environments.

Purpose of the Study:

• To develop a novel Federated Deep Learning (FDL) framework for privacy-preserving AI-driven clinical decision support.
• To enhance representation learning and model personalization in non-identically distributed (non-IID) data settings.
• To ensure robust privacy, transparency, and interpretability in AI diagnostic models.

Main Methods:

• Integration of Vision Transformers (ViT) with DINOv2 self-supervised learning for representation learning.
• Utilizing Federated Self-Supervised Learning (FedSSL) with FedProx for personalized model updates in non-IID environments.
• Implementation of differential privacy and Elliptic Curve Cryptography (ECC) for enhanced privacy and secure communication.
• Incorporation of Grad-CAM and LIME for sample-level model explainability.

Main Results:

• High diagnostic performance achieved across three medical imaging datasets: Tuberculosis (99.80% accuracy/F1-score), Diabetic Retinopathy (89.0% accuracy/F1-score), and Brain Tumors (97.1% accuracy/F1-score).
• Demonstrated efficacy in handling non-IID data through personalized federated learning.
• Validated the framework's privacy-resilience and scalability.

Conclusions:

• The proposed FDL framework effectively addresses privacy, heterogeneity, and generalizability issues in AI diagnostics.
• The system shows strong potential for real-world clinical deployment in distributed healthcare settings.
• This approach offers a scalable and privacy-resilient solution for AI-powered medical decision support.