Privacy-preserving federated learning for collaborative medical data mining in multi-institutional settings

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Imaging Analysis
• Data Privacy and Security

Background:

• Increasing reliance on AI for medical diagnostics faces critical data privacy challenges.
• Over 30% of healthcare organizations report data breaches, emphasizing the need for secure AI solutions.
• Existing methods struggle to balance model accuracy with patient data confidentiality.

Purpose of the Study:

• To investigate the integration of transfer learning and federated learning for privacy-preserving medical image classification.
• To evaluate the generalizability and robustness of the proposed framework using various AI architectures.
• To introduce and analyze a novel adaptive aggregation method for optimized federated learning.

Main Methods:

• Utilized GoogLeNet and VGG16, pre-trained on ImageNet and fine-tuned on TB X-rays, brain tumor MRIs, and diabetic retinopathy datasets.
• Employed transfer learning combined with federated learning strategies.
• Developed and analyzed a novel adaptive aggregation method dynamically switching between Federated Averaging (FedAvg) and Federated Stochastic Gradient Descent (FedSGD) based on data divergence.

Main Results:

• Achieved high classification accuracy across multiple medical imaging datasets using baseline and modern architectures (EfficientNetV2, ResNet-RS).
• Demonstrated the scalability and robustness of the proposed framework.
• The adaptive aggregation method optimized model convergence while preserving data privacy.

Conclusions:

• Transfer learning combined with federated learning provides a scalable, robust, and secure solution for medical image classification.
• The proposed framework enables healthcare institutions to train accurate AI diagnostic models without compromising patient data privacy.
• This approach addresses a critical need for secure and effective AI-driven healthcare solutions.