Federated Detection

Each participating hospital trains a local object detection model on its own annotated scans, learning to predict bounding boxes and class probabilities for pathologies. Only encrypted model updates—never images or annotations—are transmitted to the aggregation server. This preserves patient privacy while enabling the global model to learn from diverse imaging protocols and patient populations.

The global model is updated using Federated Averaging (FedAvg) applied to both the backbone feature extractor and the detection head. Local institutions train on their private data for several epochs, then send model weight updates to a central parameter server. The server computes a weighted average of these updates, with weights proportional to each institution's dataset size, ensuring equitable contribution to the global detection model.

Medical imaging datasets across hospitals are inherently non-IID (non-Independently and Identically Distributed) due to varying scanner vendors, patient demographics, and disease prevalence. Federated detection frameworks incorporate techniques to handle this heterogeneity:

• FedProx adds a proximal term to stabilize training under statistical heterogeneity
• Scaffold corrects for client drift using control variates
• Personalization layers allow site-specific fine-tuning of detection heads

To prevent membership inference attacks and model inversion that could reconstruct patient data from model updates, federated detection systems integrate differential privacy (DP). Gaussian noise is added to gradient updates before transmission, with a carefully calibrated privacy budget (ε, δ). This provides a mathematical guarantee that an adversary cannot determine whether any individual patient's scan was included in the training set.

Federated detection in healthcare typically employs a cross-silo topology rather than cross-device. A small number of reliable institutional nodes (hospitals, imaging centers) with substantial compute resources participate, as opposed to thousands of unreliable edge devices. This topology supports:

• Synchronous training rounds with guaranteed participation
• Stateful clients that maintain local model state across rounds
• Secure aggregation using trusted execution environments or multi-party computation

Object detection models like Faster R-CNN or YOLO contain millions of parameters, making raw gradient transmission bandwidth-intensive. Federated detection employs compression techniques to reduce communication overhead:

• Gradient quantization reduces 32-bit floats to 8-bit integers
• Gradient sparsification transmits only the top-k largest gradient values
• Structured updates compress weight matrices using low-rank approximations These methods reduce bandwidth by 100-1000x with minimal accuracy loss.

A specialized application of federated object detection focused on identifying suspicious regions such as pulmonary nodules, colonic polyps, or microcalcifications in mammograms. Unlike generic object detection, lesion detection models must handle extreme class imbalance—where abnormalities represent less than 0.1% of pixels—and require high sensitivity thresholds to minimize false negatives. Federated training allows these models to learn from diverse pathological presentations across institutions without centralizing protected health information.

• Key challenge: Bounding box annotation heterogeneity across radiologists
• Common architectures: RetinaNet, YOLOv8, Faster R-CNN with FPN backbones
• Validation metric: Free-response ROC (FROC) analysis at clinically relevant false-positive rates

The process of adapting a global object detection model to the specific data distribution of a local hospital's scanner vendor, acquisition protocol, or patient demographics without sharing the local target domain data. Medical imaging suffers from severe domain shift—a model trained on Siemens scanners may fail on GE equipment due to differences in reconstruction kernels, contrast timing, or voxel resolution. Federated domain adaptation techniques such as batch normalization calibration and feature alignment mitigate this without data pooling.

• Unsupervised DA: Adapts without local annotations using adversarial feature alignment
• Source-free DA: Updates model using only local unlabeled data and global model statistics
• Test-time adaptation: Adjusts normalization statistics during inference on novel scanner types

A technique for learning a common feature space or style transfer function across heterogeneous imaging scanners and protocols in a decentralized manner. Rather than adapting the detection head, harmonization transforms input images to appear as if acquired by a reference scanner, mitigating domain shift at the pixel level. Federated harmonization uses cycle-consistent generative adversarial networks trained collaboratively, where each site learns to translate between its local domain and a shared latent representation without exchanging patient scans.

• Applications: Standardizing multi-site MRI contrasts, CT kernel normalization
• Architecture: Federated CycleGAN with shared generator-discriminator pairs
• Validation: Fréchet Inception Distance (FID) computed on harmonized validation sets

The decentralized extraction and analysis of high-throughput quantitative features from medical images, enabling biomarker discovery across institutions without sharing source DICOM data. Radiomics extracts hundreds of handcrafted features—including shape descriptors, texture matrices (GLCM, GLRLM), and wavelet decompositions—from regions of interest. Federated radiomics computes these features locally and shares only aggregated statistics or model gradients, preserving patient privacy while achieving statistical power from multi-institutional cohorts.

• Feature categories: First-order statistics, shape, texture, and filter-based
• Federated feature selection: Collaborative LASSO or mutual information ranking
• Clinical use case: Imaging biomarkers for immunotherapy response prediction

The privacy-preserving training of models to classify the progression or severity of a disease from medical images, leveraging longitudinal data across institutions without centralizing patient timelines. Unlike binary detection, staging requires ordinal classification across multiple severity grades—for example, liver fibrosis stages F0-F4 or cancer TNM staging. Federated staging models must handle temporal dependencies in longitudinal scans while respecting that each institution may only see a subset of a patient's disease trajectory.

• Challenge: Label noise from inter-rater variability in staging criteria
• Approach: Federated ordinal regression with consistent grading rubrics
• Example: Multi-site diabetic retinopathy severity grading (mild, moderate, severe, proliferative)

A decentralized approach to training survival analysis or outcome prediction models directly from medical images, enabling multi-institutional prognostic biomarker development without sharing patient follow-up data. These models learn to predict time-to-event outcomes—such as overall survival or progression-free survival—from baseline imaging. Federated prognosis models must handle censored data (patients lost to follow-up) and leverage Cox proportional hazards or deep survival networks trained via federated averaging.

• Input: Baseline CT/MRI with tumor segmentation masks
• Output: Risk score or survival probability curves at 1, 3, 5 years
• Federated metric: Concordance index (C-index) aggregated across sites