Federated Learning for Multi-Center Imaging
Use this skill to design or audit a federated imaging study when raw patient data cannot be pooled. Treat federation as a distributed training architecture—not as automatic privacy, regulatory compliance, fairness, or external validation.
The approved metadata lists horizontal, vertical, split, and personalized federation together, but do not treat them as peer topology choices. Immediately decompose the design into data-ownership, computation, coordination/trust, and output-objective axes as specified below.
Core stance
- Governance before algorithms. Confirm that each site may compute and transmit the planned updates, that controller/processor roles are assigned, and that incident and withdrawal procedures exist. If this is infeasible, stop; do not solve a governance barrier with FedAvg.
- Specify orthogonal design axes. Describe data partition/ownership, computation architecture, coordination/trust, and output objective separately. Horizontal data can use full-model or split computation and can produce one global, clustered, or personalized model.
- Keep a real baseline ladder. Compare local-only models, a federated FedAvg baseline, a centralized pooled-data oracle when lawful or a clearly labeled simulation when not, and the proposed federated method.
- Apply external evaluation as a layer. After development is frozen, evaluate each eligible local, centralized, FedAvg, and proposed model on the same untouched external domain; external validation is not itself a comparator.
- Model non-IID structure explicitly. Scanner vendor, protocol, prevalence, referral pathway, annotation practice, sample size, and outcome availability can all differ by site.
- Define the adversary. Secure aggregation and differential privacy address different threats; neither protects against every leakage, poisoning, or governance failure.
- Separate collaboration from transportability. A center that trained the federation is not external validation, even if it never shared raw data.
When to open extra files
| File | Open when |
|---|---|
| references/federated-design.md | Selecting federation type, FedAvg and alternatives, site weighting, non-IID handling, communication/failure policy, or site-holdout design |
| references/privacy-governance-evaluation.md | Defining leakage threats, secure aggregation, differential privacy, controller roles, audit trails, fairness, calibration, baseline comparisons, or external-validity claims |
Workflow
1. Establish governance feasibility
Create a site-by-site feasibility table covering data-controller roles, legal basis/consent, permitted update types, cross-border restrictions, security review, model ownership, publication rights, withdrawal/deletion, incident response, and approval status. Record unresolved items as blocking decisions. Federation is not feasible until every participating site can execute the same approved protocol or an explicitly justified site-specific variant.
2. Select the federation topology
Map patient overlap, feature overlap, label location, compute capacity, network constraints, and trust boundaries. Specify four axes rather than forcing peer alternatives into one label: (1) horizontal, vertical, or hybrid data partition/ownership; (2) full-model FL, split learning, or SplitFed computation; (3) centralized/decentralized coordination and server/client trust; and (4) global, clustered/multi-task, or personalized output. Explain the selected combination—for example, horizontal data ownership + full-model synchronous coordination + honest-but-curious server + personalized heads. State entity-resolution implications for cross-site patient linkage and activation/update flows for the computation architecture.
3. Freeze the baseline ladder
Pre-specify:
- each site's local-only model;
- a pooled centralized oracle if lawful, or a clearly labeled centralized simulation;
- FedAvg as the primary federated baseline;
- the proposed heterogeneity-aware or personalized method.
Use the same eligible population, preprocessing family, architecture capacity, tuning budget, and endpoint definition wherever the comparison permits. Do not call a comparison fair when the proposed model receives more data, tuning, or compute.
4. Specify the training protocol
Lock patient-level partitions before preprocessing. Declare participating clients per round, client sampling, local epochs/steps, batch size, optimizer and learning rate, server aggregation, site weighting, number and stopping rule of communication rounds, checkpoint selection, hyperparameter-search boundary, random seeds, update compression, encryption, timeout, retry, dropout, and recovery rules. Never use the final test or held-out center to choose rounds, hyperparameters, aggregation, or personalization.
5. Write the threat model
Name protected assets, trusted components, adversaries, capabilities, attack surfaces, and acceptable residual risk. At minimum assess membership inference, gradient/update inversion, property inference, malicious-client poisoning/backdoors, model theft, collusion, and metadata leakage. Map each control—transport security, authentication, secure aggregation, clipping, differential privacy, anomaly detection, access control, and audit logging—to a named threat. For differential privacy, declare central, local, or distributed trust; clipping/noise location; who can see pre-noise individual or aggregate updates; and dropout/collusion assumptions. State whether secure aggregation prevents update-level anomaly inspection or robust aggregation and what privacy-preserving substitute is used. Report unmitigated threats rather than asserting that “data never leave the hospital.”
6. Analyze heterogeneity
Profile each site before training: population, prevalence, scanner/vendor/field strength, acquisition and reconstruction, annotation protocol, missingness, outcome ascertainment, sample size, and compute/network capability. Quantify label, feature, quantity, and concept shift. Compare FedAvg with a justified option such as FedProx, SCAFFOLD, FedBN, clustered FL, or local fine-tuning/personalization. Treat site weighting as an estimand choice, not a convenience.
7. Evaluate performance and transportability
Report patient-level metrics with uncertainty overall and by site: discrimination or task accuracy, calibration, clinically relevant thresholds/utility, failure modes, and subgroup fairness. Define whether inference is conditional on the observed sites or targets a population of new sites; name patient and site sampling units and use site-stratified, site-clustered, or hierarchical resampling accordingly. Include local-versus-federated and centralized-versus-federated paired comparisons. Use leave-one-site-out retraining or an untouched nonparticipating center to test transportability; account for dependence across leave-one-site-out folds because their training sets overlap. Label participating-site evaluation as internal, even when updates rather than raw data were shared. Treat the external set as an evaluation layer: apply it without refitting to every frozen local, centralized, FedAvg, and proposed model that can lawfully and technically be run there, and report unavailable cells rather than changing the comparator set.
8. Make the federation reproducible
Version code, containers, model initialization, preprocessing, ontology/label rules, configuration, site manifests, dependency lockfiles, seeds, and aggregation logic. Hash approved artifacts and record round-level participation, failures, update acceptance/rejection, privacy parameters, checkpoints, and deviations in a tamper-evident audit trail. Provide a rerun plan that does not require exporting patient-level data.
Output contract
Governance feasibility matrix— site approvals, roles, permitted flows, unresolved blockers.Federation topology— the selected ownership, computation, coordination/trust, and output axes, including their combined data flows and linkage assumptions.Baseline ladder— local, centralized/simulated, FedAvg, and proposed models.Training protocol— partitions, clients/round, local/server optimization, weighting, communication, stopping, failure handling, and tuning boundary.Threat-control matrix— assets, adversaries, attacks, controls, and residual risks.Heterogeneity plan— site/scanner shifts, diagnostics, method choice, and sensitivity tests.Evaluation plan— overall/site/subgroup metrics, calibration, utility, target inferential population, site/patient sampling units, dependence handling, true site holdout, and the external-evaluation matrix applied across frozen models.Reproducibility record— versions, hashes, seeds, manifests, audit events, and deviations.Bounded claims— what federation, privacy controls, and validation do and do not establish.
Quality bar
A defensible study can be reconstructed from approved artifacts, compares against local and FedAvg baselines, survives client failure without changing the estimand silently, quantifies site-level heterogeneity and calibration, and tests a genuinely unseen center. It never equates federated learning with anonymization, differential privacy, fairness, or external validity.
Handoffs
- Cohort construction, patient-level splitting, and external-validation design →
radiology-design. - CNN/Transformer architecture, optimization, and imaging input pipelines →
radiology-deep-learning. - Diagnostic metrics, confidence intervals, calibration, fairness, and model comparison →
radiology-stats. - Ethics, consent, privacy language, and governance statements →
radiology-ethics. - DICOM de-identification, data/code availability, and FAIR metadata →
radiology-data. - CLAIM/TRIPOD+AI and protocol reporting audits →
radiology-reporting. - Prospective deployment, reader studies, monitoring, and rollback →
radiology-translation.

