Federated Computed Tomography (CT)

Unlike 2D X-ray federated learning, Federated CT operates on 3D volumetric tensors—stacks of axial slices forming complete anatomical representations. Local nodes train on full DICOM series, processing spatial context across slices to detect lesions, nodules, or hemorrhages that span multiple anatomical planes. The global model must learn volumetric features without ever accessing raw voxel data, requiring specialized 3D convolutional architectures that can be efficiently aggregated across sites with heterogeneous slice thicknesses and reconstruction kernels.

Federated CT models must contend with extreme variability in radiation dose protocols across institutions. Low-dose screening scans exhibit fundamentally different noise characteristics than standard diagnostic CTs. Key challenges include:

• Noise distribution mismatch between sites using different dose modulation techniques
• Learning robust features that generalize from ultra-low-dose (<1 mSv) to standard-dose (>7 mSv) acquisitions
• Preventing the global model from overfitting to high-dose, high-contrast scans from academic centers
• Maintaining diagnostic accuracy for subtle findings like ground-glass opacities in noisy low-dose data

Each CT scanner vendor employs proprietary reconstruction algorithms—filtered back projection, iterative reconstruction, or deep learning reconstruction—that produce visually distinct image textures. Federated CT must disentangle pathology from reconstruction-induced artifacts. A nodule's apparent texture in one hospital's images may differ entirely at another site due to reconstruction kernel choices (e.g., lung vs. soft tissue kernels). Federated domain adaptation techniques are essential to learn scanner-invariant representations without sharing raw sinogram data or proprietary reconstruction parameters.

CT protocols vary in anatomical coverage—a chest CT may include the upper abdomen, while a dedicated lung screening scan excludes it. Federated CT models must handle:

• Partial anatomical fields where certain organs are inconsistently present across sites
• Variable slice ranges and z-axis coverage creating mismatched input dimensions
• The risk of the model learning spurious correlations based on scan length rather than pathology
• Registration challenges when aligning multi-phase contrast studies (arterial, venous, delayed) across institutions with different injection protocols

Multi-institutional CT datasets contain scans acquired at different contrast enhancement phases—non-contrast, arterial, portal venous, and delayed. A liver lesion appears fundamentally different in each phase. Federated CT requires explicit phase-aware training where local nodes label the contrast timing of each scan. Without centralized data pooling, coordinating phase-specific model branches becomes complex. The global aggregation must ensure that diagnostic features learned from arterial-phase scans at one site do not contaminate non-contrast model weights at another.

CT slice thickness ranges from sub-millimeter (0.625mm) for high-resolution lung imaging to 5mm for routine abdominal scans. This z-axis resolution disparity directly impacts volumetric model performance. Federated CT architectures must incorporate:

• Adaptive resampling layers that normalize voxel dimensions before feature extraction
• Multi-scale feature pyramids that remain robust across resolution variations
• Careful handling of partial volume effects that blur small structures in thick slices
• Aggregation strategies that prevent thin-slice sites from dominating gradient contributions due to higher voxel counts

A specific application of federated learning focused on training denoising and reconstruction models to maintain diagnostic accuracy in low-radiation-dose CT scans. The goal is to collaboratively learn the mapping from noisy, low-dose images to high-quality diagnostic images without requiring institutions to share their patient scans. This addresses the ALARA (As Low As Reasonably Achievable) principle by enabling the development of robust algorithms trained on diverse scanner types and patient populations, directly reducing radiation exposure risk while preserving privacy.

A specialized application of federated object detection focused on identifying suspicious regions such as pulmonary nodules, polyps, or microcalcifications across distributed medical imaging archives. This technique allows multiple hospitals to collaboratively train a highly sensitive detection model by sharing only encrypted gradient updates, not the CT scans themselves. It is critical for building robust computer-aided detection (CADe) systems that have been exposed to a diverse range of lesion presentations and rare morphologies from a global patient pool.

The collaborative optimization of inverse problem solvers for medical imaging modalities. In the context of CT, this involves learning to map raw sensor data (sinograms) to diagnostic images without aggregating raw acquisition data. Key benefits include:

• Improved generalization across different scanner geometries and vendors.
• Reduced data transfer by keeping massive raw projection datasets local.
• Accelerated innovation in deep learning reconstruction (DLR) techniques through multi-institutional collaboration.

The decentralized extraction and analysis of high-throughput quantitative features from medical images, enabling biomarker discovery across institutions without sharing the source DICOM data. Federated radiomics allows researchers to mine CT scans for textural, shape, and intensity features that correlate with tumor genotypes or patient outcomes. This overcomes the statistical power limitations of single-institution studies, unlocking the potential of imaging biomarkers for precision oncology while strictly adhering to data residency requirements.

The process of adapting a global imaging model to the specific data distribution of a local hospital's scanner or patient population without sharing the local target domain data. This is essential for CT imaging due to scanner-induced domain shift—variations in pixel spacing, reconstruction kernels, and contrast protocols across manufacturers like Siemens, GE, and Philips. Federated domain adaptation ensures a collaboratively trained lung screening model performs accurately on a specific hospital's unique imaging pipeline without centralizing sensitive scans.

A technique for training models to identify rare pathological findings or outliers in medical imaging across distributed datasets. In CT analysis, this is crucial for flagging incidental findings—unexpected abnormalities like adrenal masses or bone lesions discovered during routine chest or abdominal scans. By learning the statistical norm from diverse, multi-institutional populations without sharing images, federated anomaly detection models become highly sensitive to true outliers while reducing false positives caused by normal anatomical variation.