FedVGM: Enhancing Federated Learning Performance on Multi-Dataset Medical Images With XAI.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Deep learning advances medical imaging but faces challenges from data privacy and fragmentation.
• Existing methods often require centralizing sensitive patient data, limiting collaboration.

Purpose of the Study:

• To introduce FedVGM, a privacy-preserving federated learning framework for multi-modal medical image analysis.
• To enable collaborative diagnostics across institutions without compromising patient data privacy.

Main Methods:

• FedVGM integrates four modalities (brain MRI, breast ultrasound, chest X-ray, lung CT) using transfer learning and VGG16/MobileNetV2 ensemble.
• Evaluated three aggregation strategies, identifying median aggregation as most effective.
• Applied explainable AI techniques for clinical interpretability and validated with performance metrics and k-fold cross-validation.

Main Results:

• Achieved 97.7% ± 0.01 accuracy on the combined multi-modal dataset.
• Individual modality accuracy ranged from 91.9% to 99.1%.
• Demonstrated the effectiveness of median aggregation for model performance.

Conclusions:

• FedVGM provides a robust, scalable, and privacy-preserving solution for collaborative medical image analysis.
• The framework supports clinical deployment by ensuring data privacy and interpretability.
• Facilitates advancements in multi-modal medical diagnostics through federated learning.