Privacy-Aware Federated nnU-Net for ECG Page Digitization

TL;DR

The paper tackles privacy constraints in converting ECG page images to calibrated waveforms by implementing a privacy-aware cross-silo federated learning framework that trains a full nnU-Net backbone without sharing raw images. It combines secure aggregation with central Gaussian differential privacy, using Rényi accounting to provide auditable user-level privacy over rounds, and evaluates FedAvg, FedProx, and FedAdam under realistic non-IID page variations. Across PTB-XL derived pages, FedAdam consistently achieves faster convergence and higher late-round performance, approaching centralized baselines while preserving data locality. The work demonstrates practical deployment potential for multi-institution ECG digitization with auditable privacy guarantees and lays groundwork for future enhancements like per-user DP and personalization.

Abstract

Deep neural networks can convert ECG page images into analyzable waveforms, yet centralized training often conflicts with cross-institutional privacy and deployment constraints. A cross-silo federated digitization framework is presented that trains a full-model nnU-Net segmentation backbone without sharing images and aggregates updates across sites under realistic non-IID heterogeneity (layout, grid style, scanner profile, noise). The protocol integrates three standard server-side aggregators--FedAvg, FedProx, and FedAdam--and couples secure aggregation with central, user-level differential privacy to align utility with formal guarantees. Key features include: (i) end-to-end full-model training and synchronization across clients; (ii) secure aggregation so the server only observes a clipped, weighted sum once a participation threshold is met; (iii) central Gaussian DP with Renyi accounting applied post-aggregation for auditable user-level privacy; and (iv) a calibration-aware digitization pipeline comprising page normalization, trace segmentation, grid-leakage suppression, and vectorization to twelve-lead signals. Experiments on ECG pages rendered from PTB-XL show consistently faster convergence and higher late-round plateaus with adaptive server updates (FedAdam) relative to FedAvg and FedProx, while approaching centralized performance. The privacy mechanism maintains competitive accuracy while preventing exposure of raw images or per-client updates, yielding deployable, auditable guarantees suitable for multi-institution settings.

Figures (5)

• Figure 1: Federated training overview. The server distributes the current global nnU-Net parameters; clients train locally on page images and return updated model weights (or weight deltas) for aggregation. No raw images or signals are shared.
• Figure 2: Overview of the nnU-Net architecture and its self-configuring pipeline components used as the trainable backbone in this work.
• Figure 3: Digitization pipeline. Page pre-processing and calibration; nnU-Net trace segmentation with full-model optimization; post-processing with panel parsing; vectorization to calibrated twelve-lead signals.
• Figure 4: Global Dice over rounds on PTB-XL digitization (5-client non-IID split). Methods: FedAvgMcMahan2017FedAvg, FedProxLi2020FedProx, FedAdamReddi2021FedOpt. Dashed line: centralized reference with the same backbone.
• Figure 5: Stacked qualitative examples from the three federated aggregation methods on held-out ECG pages. Consistent with the quantitative results, FedAdam produces cleaner masks with fewer gaps at grid crossings and smoother centerlines after vectorization.