Federated Learning for Medical Image Classification
Privacy-preserving learning across distributed cervical cell classification datasets. Research comparing centralized and federated approaches under realistic non-IID conditions with client-level explainability analysis.
Challenge
Medical data is often legally or practically impossible to centralize. Hospitals, clinics, and research institutions hold valuable datasets that cannot be combined due to:
- Patient privacy regulations (HIPAA, GDPR, local data protection laws)
- Institutional data governance policies
- Technical infrastructure limitations
- Cross-border data transfer restrictions
The research investigates whether distributed datasets can contribute to shared learning while remaining in place. Focus on cervical cell classification, a real medical imaging task with clinical relevance and available benchmark datasets.
Technical Approach
Centralized Baseline
Established centralized training baseline using CNN and transformer architectures on pooled data. This upper-bound performance provides the comparison target for federated methods.
Federated Learning Algorithms
Implemented and compared five federated learning approaches:
- FedAvg: Foundational federated averaging algorithm
- FedProx: Addresses client heterogeneity with proximal term
- FedBN: Maintains local batch normalization statistics
- SCAFFOLD: Variance reduction through control variates
- MOON: Model-contrastive learning for consistency
Non-IID Data Distribution
Real federated settings have non-identically distributed (non-IID) data across clients. Simulated realistic scenarios where different clients have different class distributions, sample sizes, and data characteristics. Tested algorithms under both IID and non-IID conditions.
Architecture Comparison
Evaluated CNN and transformer-based architectures in the federated setting. Assessed how architectural choices interact with federated learning algorithms and non-IID data distributions.
Client-Level Explainability
Applied explainability methods at the client level to understand how local data characteristics and federated training affect model behavior. Essential for medical AI where understanding decision-making is critical.
Findings
The research provides structured comparison of how centralized and federated approaches perform across realistic non-IID client conditions. Identified which algorithms handle data heterogeneity effectively and how performance degrades compared to centralized training.
Technologies & Methods
- Algorithms: FedAvg, FedProx, FedBN, SCAFFOLD, MOON
- Architectures: CNN baselines, transformer models
- Data: Cervical cell classification with IID and non-IID partitioning
- Evaluation: Centralized baseline comparison, client-level metrics
- Explainability: Client-level XAI analysis
- Framework: PyTorch with federated learning extensions
Implications for Medical AI
- Privacy-preserving learning is feasible: Federated learning enables model training across distributed medical data without centralization, addressing major regulatory and institutional barriers.
- Non-IID matters: Real-world federated settings have heterogeneous data. Algorithm selection must account for how methods handle client distribution differences.
- Trade-offs are real: Federated learning typically achieves lower performance than centralized training. Quantifying this gap helps decision-makers assess whether privacy benefits justify accuracy cost.
- Explainability remains possible: Client-level XAI methods provide transparency even in federated settings, essential for clinical adoption.
- Algorithm choice is context-dependent: No single federated learning algorithm dominates all scenarios. Selection depends on data characteristics, infrastructure constraints, and privacy requirements.
Discuss Privacy-Preserving AI
If you have distributed medical data, regulatory constraints preventing centralization, or multi-institutional AI requirements, we can assess federated learning feasibility and propose an implementation strategy.
Start a Conversation →