Domain Adaptation

The fundamental problem that domain adaptation solves. Domain shift is the statistical mismatch between the source distribution (e.g., natural images from ImageNet) and the target distribution (e.g., CT scans from a specific hospital). In medical imaging, this arises from:

• Scanner variability: Different manufacturers (Siemens vs. GE) produce different pixel intensity distributions
• Protocol divergence: Varying acquisition parameters (slice thickness, contrast agents, reconstruction kernels)
• Population drift: Demographic differences between training and deployment patient cohorts

Without adaptation, a model trained on one hospital's data can experience 20-40% performance degradation when deployed at another site.

A neural network architecture that learns domain-invariant features through adversarial optimization. The system consists of three components:

• Feature extractor: A CNN or Vision Transformer that encodes input images into a latent representation
• Label predictor: A classifier trained to predict the diagnostic label from the features
• Domain classifier: A discriminator trained to identify which domain (source or target) the features came from

The key innovation is the gradient reversal layer placed between the feature extractor and domain classifier. During backpropagation, gradients from the domain classifier are multiplied by a negative constant, forcing the feature extractor to produce representations that actively confuse the domain discriminator. This yields features that are discriminative for the task but invariant to domain.

A more ambitious objective than domain adaptation. While adaptation assumes access to unlabeled target domain data during training or testing, domain generalization requires the model to perform well on completely unseen target domains without any exposure. Strategies include:

• Multi-source training: Training on diverse domains to learn robust, shared representations
• Data augmentation: Aggressive transformations (style transfer, intensity perturbations) to simulate domain shift during training
• Meta-learning: Training across domain episodes to learn how to generalize
• Invariant risk minimization: Learning representations whose relationship to labels remains stable across environments

In medical imaging, domain generalization is critical for regulatory approval, as a diagnostic model must work reliably on scanners and populations it was never trained on.

A lightweight alternative to full domain adaptation that updates a model during inference on the target domain. Rather than requiring a separate training phase, test-time adaptation adjusts the model on-the-fly using only the current test sample or a small batch. Common approaches:

• Batch Normalization recalibration: Updating the running mean and variance statistics of BN layers using target domain statistics. This alone can recover significant performance without any gradient updates.
• Entropy minimization: Fine-tuning the model to produce confident predictions on target samples by minimizing prediction entropy
• Rotation prediction: Using a self-supervised auxiliary task (predicting image rotation) to adapt feature extractors to the target domain

Test-time adaptation is particularly valuable in point-of-care deployment where a model must immediately adapt to a new scanner without offline retraining.

An unpaired image-to-image translation technique that transforms source domain images to appear as if they originated from the target domain. Unlike paired translation (which requires corresponding images in both domains), CycleGAN learns a mapping using only unpaired collections from each domain. The architecture uses:

• Two generators: One mapping source→target, another mapping target→source
• Two discriminators: One for each domain, distinguishing real from translated images
• Cycle consistency loss: Enforcing that translating an image to the target domain and back recovers the original

In medical imaging, CycleGAN can harmonize CT scans across scanners or convert between modalities (e.g., MR to CT synthesis). The translated images can then be used to train a task model that generalizes to the target domain.

A specialized preprocessing pipeline designed to standardize medical images across acquisition devices before model training. Unlike learned adaptation, harmonization applies deterministic transformations to reduce domain shift at the pixel level:

• Hounsfield Unit normalization: Rescaling CT intensities to standardized HU values based on tissue-specific references (air = -1000, water = 0)
• Histogram matching: Aligning the intensity distribution of target images to match a reference scanner's histogram
• Z-score normalization: Standardizing intensities to zero mean and unit variance per scan
• Bias field correction: Removing low-frequency intensity non-uniformities caused by MR scanner imperfections

Harmonization is often a prerequisite for domain adaptation, reducing the magnitude of shift that the learning algorithm must overcome.