Federated Low-Dose CT

The core mechanism involves training a deep convolutional neural network to map noisy low-dose CT images to their high-dose diagnostic equivalents. Each hospital trains locally on its own paired low-dose/high-dose datasets, computes model weight updates, and transmits only these encrypted gradients to a central aggregation server. The raw sinogram data and reconstructed DICOM images never leave the hospital's firewall. This paradigm leverages Federated Averaging (FedAvg) to combine updates into a global model that captures diverse scanner geometries and patient demographics without ever seeing the underlying data.

Clinical low-dose protocols typically aim for a 70-90% reduction in radiation exposure compared to standard diagnostic scans. The federated model is trained to recover the diagnostic quality lost by this aggressive dose reduction. Key metrics include:

• Peak Signal-to-Noise Ratio (PSNR) improvement over filtered back projection
• Structural Similarity Index (SSIM) preservation relative to full-dose ground truth
• Lesion conspicuity maintained for radiologist assessment
• Contrast-to-noise ratio enhancement in soft tissue regions The goal is to achieve diagnostic equivalence with sub-milliSievert effective doses, particularly critical for pediatric and screening populations where cumulative radiation risk is a primary concern.

CT data across institutions is inherently non-independent and identically distributed (non-IID) due to:

• Scanner vendor variability: Siemens, GE, Philips, Canon systems produce different noise textures and reconstruction kernels
• Protocol heterogeneity: Tube voltage (kVp), tube current (mAs), pitch, and slice thickness vary by site
• Patient population skew: Demographics, body habitus, and pathology prevalence differ regionally Federated Low-Dose CT frameworks must incorporate domain adaptation layers or personalized federated learning techniques to prevent the global model from overfitting to dominant sites. Techniques like FedProx add proximal terms to local objectives, stabilizing convergence across heterogeneous nodes.

Even gradient updates can leak information about individual patients through gradient inversion attacks. Federated Low-Dose CT implementations integrate differential privacy (DP) mechanisms to provide formal mathematical guarantees:

• Gaussian noise injection into gradients before transmission
• Gradient clipping to bound individual contribution sensitivity
• Privacy budget (ε, δ) tracking across training rounds
• Secure aggregation protocols that prevent the central server from inspecting individual updates A typical privacy budget of ε < 8 provides meaningful protection while maintaining clinically acceptable denoising performance. This is essential for HIPAA compliance and cross-border collaborations under GDPR.

CT denoising models, often based on U-Net, ResNet, or vision transformer backbones, contain millions of parameters. Transmitting full model updates each round is bandwidth-intensive. Optimization strategies include:

• Gradient compression via sparsification or quantization to 8-bit integers
• Federated distillation where only soft labels on a public proxy dataset are exchanged
• Split learning configurations where the denoising network is partitioned, with early layers local and only activations shared
• Asynchronous update protocols that allow straggler sites to participate without blocking faster nodes These techniques reduce communication overhead by 100-1000x while preserving model convergence.

A critical evaluation step is testing the federated model on held-out scanner types and dose levels not seen during training. This validates true generalization rather than memorization of site-specific noise patterns. Radiologist reader studies are often conducted where:

• Federated denoised images are compared against standard low-dose and full-dose reconstructions
• Diagnostic accuracy for specific pathologies (e.g., pulmonary nodules, liver lesions) is measured
• Inter-reader agreement is assessed using Cohen's kappa statistics The ultimate benchmark is non-inferiority to full-dose diagnostic performance, ensuring the federated model is clinically deployable across diverse, unseen hospital environments.

Low-dose CT scans from different manufacturers exhibit domain shift due to varying beam-hardening corrections, iterative reconstruction algorithms, and detector efficiencies. Federated Image Harmonization trains a style transfer or domain adaptation network across sites to standardize image appearance without sharing patient data. This is critical for Federated Low-Dose CT because a denoising model trained on Siemens scanners may catastrophically fail on GE scanners. Techniques include federated adversarial training with site-specific discriminators and feature distribution matching via maximum mean discrepancy.

The fundamental statistical challenge undermining Federated Low-Dose CT. Patient demographics, scan protocols, and disease prevalence vary wildly across hospitals, violating the independent and identically distributed assumption of standard optimization. A model trained on predominantly lung cancer screening scans from a tertiary center will perform poorly on trauma scans from a rural hospital. Mitigation strategies include FedProx (adding a proximal term to local objectives), SCAFFOLD (correcting for client drift via control variates), and personalized federated learning layers that allow site-specific fine-tuning of the final denoising blocks.