Federated Multi-Modal AI for Uterine Cancer: A Practical Framework for Classification and Clinical Insight

This diagram presents a comprehensive and practical framework for uterine cancer classification using multi-modal data and federated learning, enriched with explainable AI (XAI) mechanisms. It integrates histopathology (WSI), MRI, and clinical tabular data into a unified pipeline that respects privacy, enhances interpretability, and supports real-world deployment.

The data pipeline begins with preprocessing tailored to each modality. WSIs undergo patch extraction, stain normalization, and blur detection to ensure visual quality. MRI data is registered to a reference sequence, intensity-normalized, and optionally segmented to isolate regions of interest. Tabular data is imputed, scaled, and flagged for protected attributes to support fairness audits.

Each modality is processed through a specialized backbone: MIL for histology patches, 3D CNN for MRI volumes, and MLP for tabular features. These embeddings are fused using late fusion, gated fusion, or co-attention strategies, allowing flexible integration based on data quality and clinical relevance. Explainability modules—Grad-CAM, SHAP, and attention heatmaps—are embedded throughout, enabling both global and local interpretation of model decisions.

Federated learning is structured across cross-silo hospital sites, using FedAvg for aggregation and optional personalization via local heads. Privacy is preserved through differential privacy (DP-SGD), secure aggregation, and strict control over personally identifiable information. Communication protocols include gradient compression and retry logic to ensure robustness.

Training objectives combine classification loss with consistency and fairness regularizers. Evaluation protocols include per-site stratified splits, leave-one-site-out validation, and temporal robustness checks. Explainability is operationalized through case cards, counterfactual probes, and bias audits, ensuring transparency and clinical trust.

Reproducibility is enforced via modular pipelines, config tracking, and ethical audit hooks. Deployment pathways include integration with WSI viewers, PACS, and EHR systems, with on-prem inference and post-deployment monitoring. Human factors are addressed through escalation rules and clinician training on explanation literacy.

This framework balances technical rigor with clinical usability, offering a scalable, interpretable, and privacy-preserving solution for uterine cancer classification.