FedMed — Federated Generative AI for Chest X-Ray Diagnosis
Designed a federated learning system letting multiple hospitals train one model together without any hospital's data leaving its own servers, and wrote the full technical report on the architecture.
Technologies
Engineering Focus
Engineered through the complete development lifecycle: requirements analysis, database normalization (3NF), modular service architecture, concurrency control, and systematic testing.
01. The Real-World Problem Space
Medical diagnostic models demand large multi-center radiology datasets to generalize, yet healthcare privacy regulations strictly forbid transmitting raw patient records outside hospital servers.
02. Strategic Engineering Approach
Designed a decentralized federated learning architecture where independent clinical nodes compute parameter gradients on local hardware and transmit only aggregated updates.
System Architecture & Service Flow
Local Hospital Silos → On-Premise Deep Learning → Encrypted Gradient Aggregation → Global Model Update
Decentralized Hospital Nodes
Simulates partitioned radiology environments with private, non-shared patient chest X-ray datasets.
On-Premise Model Training
Hospital nodes train local PyTorch neural networks on private scans without external data transmission.
Federated Parameter Aggregation
Applies Federated Averaging (FedAvg) algorithms to combine model weights with differential privacy guarantees.
Consortium Diagnostic Distribution
Distributes the aggregated global model back to all participating medical centers for accurate clinical inference.
Technical Challenges Overcome
- •Mitigating divergence caused by non-IID (heterogeneous) data distributions across different scanner manufacturers.
- •Preserving computational efficiency while ensuring strict mathematical privacy guarantees.
Demonstrated Outcome & Verification
Wrote a comprehensive technical report demonstrating robust multi-institutional diagnostic convergence with zero patient data leakage.
Future Roadmap
Integrating zero-knowledge proofs and secure multi-party computation (SMPC) for trustless cross-institutional verification.