Federated Domain Adaptation

A technique that pits a domain discriminator against a feature extractor in a minimax game. The global model learns to produce domain-invariant feature representations by confusing the discriminator, which tries to identify the source institution of the data. This alignment occurs without sharing raw images, as only gradients or activations are exchanged.

• Mechanism: Gradient Reversal Layer (GRL) flips gradients during backpropagation
• Benefit: Mitigates scanner-specific biases (e.g., GE vs. Siemens MRI)
• Challenge: Requires careful tuning of the adversarial weight to avoid mode collapse

A lightweight strategy that avoids sharing the running mean and variance statistics of Batch Normalization layers. Each client maintains its own local BN parameters, which capture domain-specific style information, while sharing only the convolutional and linear layer weights with the server.

• Mechanism: Local BN statistics act as a domain-specific normalizer
• Benefit: Simple to implement; negligible communication overhead
• Example: Effective for harmonizing H&E stain variation in digital pathology

Instead of adapting to a specific target domain, this approach trains a model to be robust to unseen domains by design. It leverages the diversity of source domains across institutions to learn invariant causal features, often using meta-learning or distributionally robust optimization within the federated aggregation step.

• Mechanism: Episodic training simulating domain shift during local updates
• Benefit: Zero-shot adaptation to new hospitals joining the network
• Challenge: Requires significant domain diversity among participating clients

A semi-supervised approach where a global model generates pseudo-labels for a local hospital's unlabeled target domain data. The local model is then fine-tuned on these high-confidence predictions, iteratively adapting to the local distribution. Only the refined model weights are shared back.

• Mechanism: Teacher-student paradigm with confidence thresholding
• Benefit: Leverages abundant unlabeled local data
• Risk: Confirmation bias if pseudo-labels are noisy; requires robust filtering

Instead of sharing raw features or images, clients exchange class prototypes—compact statistical summaries of feature representations for each category. Local models are regularized to minimize the distance between their local prototypes and the global prototypes, aligning the feature space without exposing data.

• Mechanism: L2 or cosine distance minimization between local and global prototypes
• Benefit: Highly communication-efficient and privacy-preserving
• Example: Aligning 'tumor' vs. 'normal' feature centroids across radiology departments

A paradigm where a pre-trained federated model adapts on-the-fly to a single patient's scan during inference, without any prior access to the local dataset. This is critical for handling unexpected scanner malfunctions or rare acquisition artifacts in real-time clinical workflows.

• Mechanism: Entropy minimization or rotation prediction on the test sample itself
• Benefit: No local training dataset required; instant personalization
• Constraint: Must be computationally light to run on edge devices or PACS workstations

A preprocessing sibling to domain adaptation that learns to standardize image appearance across sites without sharing raw data. While adaptation modifies the model, harmonization transforms the images themselves. Key approaches:

• CycleGAN-based style transfer in federated settings
• Contrast normalization using shared reference distributions
• Latent space alignment to create scanner-agnostic representations This is critical for quantitative radiomics where pixel intensity values must be comparable across institutions.