Federated Denoising

Leverages the Noise2Noise paradigm where models learn to denoise images using only pairs of independently noisy observations of the same underlying signal, eliminating the need for clean ground truth. In a federated setting, each institution trains locally on its own noisy image pairs—such as quarter-dose vs. half-dose CT reconstructions—and shares only model gradients. The global model learns a universal denoising function without ever accessing raw patient scans or requiring pristine reference images.

Addresses the fundamental challenge of scanner-induced domain shift in medical imaging. A denoising model trained solely on Siemens MRI data may fail on GE scanners due to different noise characteristics. Federated denoising exposes the global model to diverse noise distributions—Gaussian, Rician, and non-central Chi—across vendors, field strengths, and acquisition protocols. The aggregated model learns scanner-agnostic denoising features, achieving robust performance on unseen hardware without requiring centralized data harmonization.

Enables radiology departments to collaboratively push the boundaries of ALARA (As Low As Reasonably Achievable) radiation dosing. By training denoising models across institutions, the global model learns to recover diagnostic-quality images from ultra-low-dose scans—potentially sub-milliSievert levels—without any institution exposing its patient dose protocols or images. This is critical for pediatric imaging and longitudinal screening programs where cumulative radiation exposure must be minimized while maintaining diagnostic accuracy.

Goes beyond pixel-wise metrics like Mean Squared Error (MSE) by incorporating perceptual loss functions computed in feature space. Local nodes compute perceptual losses using shared VGG or ResNet backbones pre-trained on public datasets, comparing feature activations between denoised outputs and reference images. The federated aggregation of these perceptual gradients ensures the global model preserves diagnostically relevant textures and edge details—critical for radiologist acceptance—rather than producing overly smoothed, visually plausible but clinically misleading results.

Implements noise-level map conditioning where the denoising network accepts an estimated noise standard deviation as an additional input channel. In federated training, each site computes local noise estimates using wavelet-based or patch-based methods on their specific scanner data. The global model learns to adapt its denoising strength dynamically based on the input noise level, enabling a single deployed model to handle variable dose protocols—from standard to ultra-low-dose—without retraining or manual parameter tuning.

Protects the federated denoising model from poisoned or corrupted updates originating from malfunctioning scanners, misconfigured reconstruction pipelines, or adversarial actors. Employs robust aggregation rules such as Krum, trimmed mean, or median-based fusion at the central server to discard outlier gradients. This ensures that a single institution's faulty noise model—perhaps due to a miscalibrated CT detector—does not degrade the global denoising performance across all participating hospitals.

A mathematical guarantee that prevents the reconstruction of individual patient scans from the model updates shared during federated denoising. By adding calibrated statistical noise to gradients, it bounds the privacy loss (ε), ensuring that a local low-dose CT scan cannot be inferred from the global model. This is critical for satisfying HIPAA and GDPR requirements in multi-institutional studies.

The primary challenge in federated denoising. Noise artifacts are not identically distributed across hospitals due to scanner vendor variability, acquisition protocols, and patient demographics. A model trained on Siemens low-dose CT noise may fail on GE scanner noise. Techniques like FedProx and personalized federated learning address this statistical heterogeneity by allowing local models to diverge from the global consensus.

The foundational aggregation algorithm for federated denoising. Each hospital trains a local denoising model (e.g., a U-Net or ResNet) on its noisy-clean image pairs, then sends only the model weights to a central server. The server computes a weighted average of these updates to produce a global model. This simple yet effective protocol eliminates the need to centralize raw DICOM data.

A cryptographic technique that allows the central server to perform the weighted averaging of model updates while they remain encrypted. In the context of federated denoising, this prevents even the aggregator from inspecting individual hospital contributions. Combined with secure multi-party computation (SMPC), it provides defense-in-depth against honest-but-curious servers.

A class of neural architectures designed to remove noise without prior knowledge of the specific noise level or type. In a federated setting, blind denoising is essential because the global model must generalize to unseen noise distributions from new participating hospitals without requiring them to disclose their scanner characteristics. Architectures like DnCNN and CBDNet are common baselines.

A technique to bridge the gap between the global denoising model and a local hospital's specific data distribution. By aligning feature representations without sharing images, the model adapts to local noise patterns. This is often implemented via adversarial training or feature distribution matching at each client node, ensuring diagnostic quality is maintained across heterogeneous scanners.