Federated Radiomics

Radiomic features—including first-order statistics, shape descriptors, and texture matrices (GLCM, GLRLM, GLSZM)—are computed locally at each institution. Only the derived quantitative features, not the original DICOM images, are shared with the central aggregation server. This preserves patient privacy while enabling collaborative model training across geographically distributed hospital networks.

A critical challenge in federated radiomics is the batch effect introduced by different scanner vendors, acquisition protocols, and reconstruction kernels. Techniques like ComBat harmonization are applied in a federated manner to normalize feature distributions across sites without requiring raw pixel data to leave the local institution. This ensures that the global model learns biological signal rather than scanner-specific artifacts.

Radiomic pipelines often extract thousands of features, many of which are redundant or non-informative. In a federated setting, distributed feature selection algorithms identify the most predictive and stable features across all sites. Methods include federated versions of LASSO regularization, mutual information maximization, and recursive feature elimination, ensuring the final model is both parsimonious and generalizable.

The final radiomic signature—a weighted combination of selected features—is constructed without any single institution seeing another's patient-level data. Secure aggregation protocols combine locally computed model updates, while differential privacy mechanisms add calibrated noise to gradient updates. This guarantees that even if the aggregator is compromised, individual patient radiomic profiles cannot be reconstructed.

Federated radiomics enables external validation at scale. A model trained across multiple sites can be evaluated on held-out cohorts at each institution without data centralization. This provides robust estimates of AUC, sensitivity, and specificity across diverse patient populations, imaging protocols, and disease prevalences—critical for regulatory approval and clinical translation.

Radiomic features serve as the imaging component in broader federated multi-modal fusion architectures. Quantitative image descriptors are combined with genomic data (radiogenomics), electronic health records, and pathology reports within a joint federated learning framework. This enables holistic patient modeling where imaging biomarkers are contextualized by molecular and clinical data across institutional boundaries.

The computational process of converting medical images into mineable high-dimensional data. Algorithms quantify tumor phenotype by extracting features describing:

• First-order statistics: Histogram-based metrics (mean, skewness, kurtosis) of voxel intensities
• Shape features: Compactness, sphericity, and surface-to-volume ratio of segmented regions
• Texture matrices: Gray-Level Co-occurrence Matrix (GLCM) and Gray-Level Run-Length Matrix (GLRLM) capturing spatial relationships
• Wavelet features: Decompositions of images into frequency sub-bands to capture multi-scale patterns

In a federated context, feature extraction occurs locally at each institution using standardized algorithms, ensuring only derived quantitative values—not images—are shared.

A critical preprocessing step addressing scanner-induced variability in radiomic features across institutions. Without harmonization, features from different CT or MRI scanners are non-comparable.

Key techniques include:

• ComBat harmonization: Originally developed for genomics batch effects, adapted to radiomics to remove scanner-specific noise while preserving biological signal
• Distributed ComBat: A federated variant where each site computes local moments, shares only aggregate statistics, and applies a global harmonization transform
• Deep learning normalization: Using domain-adversarial networks to learn scanner-invariant feature representations

Harmonization is essential for building robust federated models that generalize across heterogeneous imaging equipment.

The process of identifying the most predictive and stable radiomic features across distributed datasets without centralizing data. This step combats the curse of dimensionality inherent in high-throughput radiomics.

Approaches include:

• Federated mutual information: Computing feature-target dependency scores locally and aggregating them via secure summation
• Stability selection: Identifying features consistently selected across bootstrap samples at each site, then federating the stability scores
• LASSO with distributed optimization: Using alternating direction method of multipliers (ADMM) to solve sparse regression across nodes
• Variance thresholding: Eliminating near-zero-variance features using federated computation of variance statistics

Effective feature selection reduces overfitting and improves model interpretability for clinical adoption.

The rigorous process of confirming that a radiomic biomarker generalizes across independent, unseen institutional cohorts. Unlike centralized validation, federated validation never pools test data.

Validation workflow:

• Internal-external validation: Each site serves as the test set once while others train, rotating through all institutions
• Federated AUC computation: True positive and false positive rates are computed locally and aggregated to construct a global ROC curve
• Calibration assessment: Expected vs. observed outcome probabilities are compared using federated Brier score calculation
• Decision curve analysis: Evaluating net clinical benefit across threshold probabilities without sharing patient-level predictions

This process provides the evidence required for regulatory approval of imaging biomarkers.

The analysis of temporal changes in radiomic features extracted from longitudinal imaging studies—e.g., pre- and post-treatment CT scans. Delta-radiomics captures treatment response dynamics invisible to single timepoint analysis.

Federated challenges include:

• Intra-patient alignment: Deformable registration of serial scans must occur locally before feature extraction
• Temporal harmonization: Accounting for protocol changes or scanner upgrades between timepoints at each institution
• Paired statistical testing: Federated computation of paired differences and their statistical significance across distributed cohorts
• Survival delta-modeling: Building Cox proportional hazards models on feature change vectors using federated survival analysis

Delta-radiomics is particularly valuable for adaptive radiotherapy and neoadjuvant chemotherapy response assessment.

Integrated predictive models that combine quantitative imaging features with clinical variables (age, lab values, staging) in a federated framework. Nomograms provide individualized risk scores for clinical decision-making.

Construction involves:

• Federated Cox regression: Building survival models that incorporate both radiomic signatures and clinical covariates using distributed optimization
• Variable contribution weighting: Computing each predictor's relative importance through federated permutation importance or SHAP values
• Calibration across subgroups: Ensuring nomogram accuracy across demographic and clinical subgroups using federated fairness metrics
• Deployment-ready scoring: Converting model coefficients into point-based systems that clinicians can use at the bedside

These nomograms bridge the gap between complex machine learning outputs and actionable clinical tools.