Privacy-Aware Federated nnU-Net for ECG Page Digitization
Nader Nemat 1,∗ \orcid 0000-0003-2548-6740 · 2025
Read Privacy-Aware Federated nnU-Net for ECG Page Digitization by Nader Nemat 1,∗ \orcid 0000-0003-2548-6740 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-22387.txt.
Introduction
Nader Nemati presents a privacy aware way to turn paper ECG pages into digital signals without pooling patient images in one place. Training happens across sites, each keeping its own scans, while a shared segmentation model learns from all of them. Sites differ in layout, grid style, scanner traits, and noise, which makes joint training hard.
The setup tests three rules for merging site updates and adds secure aggregation plus user level noise for privacy. On ECG pages made from a public data set, the adaptive merging rule converged faster and came close to central training. Raw images and single site updates stayed hidden.
Catalog record
| CREATOR | Nader Nemat 1,∗ \orcid 0000-0003-2548-6740 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-22387
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Privacy-Aware Federated nnU-Net for ECG Page Digitization Nader Nemat 1,∗ \orcid 0000-0003-2548-6740 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 Rényi 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. keywords ECG digitization, federated learning, privacy-preserving ML, privacy-aware, image-to-signal reconstruction, clinical AI, digitization 1 Introduction Electrocardiogram (ECG) page images remain pervasive in clinical archives and day-to-day workflows, while most analytical pipelines assume access to digitally sampled waveforms. Public corpora such as PTB-XL demonstrate the scientific and translational value of curated digital signals for benchmarking and model development. Nevertheless, many health systems retain decades of paper or scanned ECGs that are costly to query and prone to loss [ 32 ] . Converting page images into calibrated waveforms preserves longitudinal clinical history, enables secondary analyses at scale, and supports interoperable storage and retrieval across institutions. Recent work has shown that deep learning can recover high-fidelity traces from printed pages by segmenting the trace, suppressing grid artifacts, and reconstructing lead-wise signals with strong agreement to digital ground truth [ 1 , 19 ] . The 2024 George B. Moody PhysioNet Challenge further catalyzed progress on image-to-signal reconstruction and image-based classification, consolidating best practices for layout-aware post-processing and calibration-aware vectorization [ 3 , 4 , 5 ] . Centralized training remains the norm for these pipelines but is often infeasible across institutions due to regulatory, governance, and operational constraints. Federated learning (FL) offers a natural alternative by training across sites without moving raw images, with canonical aggregators such as Federated Averaging (FedAvg), proximal regularization (FedProx), and adaptive server optimization in the FedOpt family (FedAdam), addressing non-IID data and client heterogeneity [ 6 , 8 , 9 ] . In medical imaging and cardiology, cross-site FL has approached centralized performance while preserving data locality, motivating its use for ECG digitization [ 10 ] . More broadly, deep learning on sequential data has proven effective outside biomedicine as well; for example, combining CNNs trained at multiple temporal resolutions improves forecasting performance on financial time series [ 7 ] . Yet, evidence specific to image-to-signal ECG digitization under federated constraints remains limited compared with time-series classification and echocardiography modeling, leaving open questions about optimization dynamics, utility–privacy trade-offs, and communication efficiency in this setting [ 11 ] . We adopt a privacy-preserving cross-silo FL design that matches clinical requirements and our optimization pipeline. At each round, clients compute full-model updates on local pages and participate in secure aggregation (SecAgg) so that the server can only recover a masked sum of clipped updates once a participation threshold is met—individual updates remain hidden via pairwise one-time masks that cancel in aggregate [ 14 ] . On the server, we enforce central user-level DP by adding Gaussian noise to the aggregated, clipped update and composing privacy loss across rounds with a Rényi moments accountant [ 15 , 21 ] . Compared with local DP that perturbs each client’s update, central DP applied after summation achieves better utility at a fixed privacy target because effective per-client noise scales down with the cohort size; SecAgg alone, while concealing single-site updates, does not bound inference from model histories, hence the combination of SecAgg and central DP, achieves auditable guarantees on the released sequence of aggregates. This mechanism is a drop-in at aggregation time and requires no changes to client-side learning beyond norm clipping, making it compatible with full-model nnU-Net training and standard aggregators, FedAvg, FedProx, and FedAdam, used in this work [ 6 , 8 , 9 ] . 2 Related works ECG image digitization has progressed from rule-based extraction to fully learned, segmentation-to-vectorization pipelines with high agreement to digital ground truth [ 1 ] . Recent systems pair robust page normalization and thin-structure segmentation with calibration-aware vectorization, and community efforts around the 2024 George B. Moody PhysioNet Challenge focused the task and released stronger baselines and artifacts for benchmarking [ 19 , 3 , 4 , 5 ] . Within these pipelines, nnU - Net is a frequent backbone due to its self-configuration and strong biomedical-segmentation performance [ 20 ] . Challenge materials and contemporaneous datasets emphasize realistic renderings, scanner artifacts, and layout variability to stress-test reconstruction, consistent with broader medical-imaging work that reports strong centralized baselines but growing interest in distributed training when governance constrains data pooling [ 10 ] . Federated learning (FL) offers an alternative to centralization under cross-site heterogeneity. Canonical optimizers include sample-size–weighted Federated Averaging (FedAvg), proximal regularization with FedProx, and adaptive server methods in the FedOpt family (FedAdam), each mitigating client drift to different degrees [ 6 , 8 , 9 ] . To address privacy, secure aggregation (SecAgg) prevents the server from inspecting any single client update by revealing only a masked sum after a participation threshold [ 14 ] . Because SecAgg alone does not bound inferences from the released aggregates or model history, many deployments combine it with central DP, adding calibrated Gaussian noise to the post-aggregation vector and composing privacy loss across rounds via Rényi accounting [ 15 , 21 ] . Compared with local DP that perturbs each client update, central DP applied after summation typically achieves better utility at a fixed privacy target because the effective per-client perturbation scales as 1/\sqrt{|\mathcal{S}|} with the cohort size |\mathcal{S}| , while remaining lightweight relative to heavier HE/MPC/TEE pipelines; these patterns align with cross-silo imaging FL where full-model synchronization is common [ 14 ] . 3 Materials & Methods 3.1 Data The PTB-XL dataset is utilized as the authoritative source of twelve-lead ECG waveforms. PTB-XL contains 21,837 clinical 12-lead ECG records (10 s) from 18,885 patients with waveform files and richly curated metadata, including SCP-ECG labels and basic demographics. Signals are provided at 500\text{\,}\mathrm{H}\mathrm{z} and 100\text{\,}\mathrm{H}\mathrm{z} sampling rates. These properties make PTB-XL well-suited for constructing paired page–signal examples and for benchmarking digitization under realistic diagnostic diversity. This study focuses on developing and analysing federated learning methodology on nn-Unet deep neural network model over the digitization of ECG images which specifies ECG image formats, WFDB headers, and the target taxonomy for image-based algorithms. That framing emphasizes that ECG images which include synthetic renderings from digital signals realized by creases, shadows, and faded ink, and requires methods to be robust across this spectrum [