Domain Adaptation

The most direct adaptation approach: take a pre-trained source model and continue training on a small labeled dataset from the target domain. In medical imaging, this often means annotating just 50–100 scans from the new scanner to recalibrate feature extractors. Key considerations:

• Use a lower learning rate (typically 1/10th of original) to prevent catastrophic forgetting
• Freeze early layers that capture universal features (edges, textures) and only fine-tune deeper, domain-specific layers
• Risk: overfitting to the small target set if not combined with strong regularization

A technique that pits a domain discriminator against the feature extractor in a minimax game. The feature extractor learns to produce representations that are indistinguishable between source and target domains, while the discriminator tries to identify which domain a feature vector came from. Architecture:

• A gradient reversal layer flips the sign of gradients during backpropagation, enabling end-to-end adversarial training
• Particularly effective for scanner-to-scanner adaptation where texture and contrast variations dominate
• Does not require any target domain labels, making it a form of unsupervised domain adaptation

Uses unpaired image-to-image translation to stylistically convert source-domain images to appear as if they came from the target domain (or vice versa). A CycleGAN learns a bidirectional mapping with cycle-consistency loss, ensuring that translating an image to the target style and back recovers the original. Medical imaging applications:

• Converting contrast-enhanced CT to non-contrast CT appearance
• Normalizing H&E stained pathology slides across different lab protocols
• Adapting MRI scans across different field strengths (1.5T vs 3T)
• The detection model is then trained or tested on the style-normalized images

Instead of adapting to a specific target domain, this technique exposes the model to extreme variation during training so that the target domain appears as just another variation. Applied aggressively in sim-to-real transfer for surgical tool detection. Implementation:

• Randomize image properties: brightness, contrast, noise levels, blur, rotation, and elastic deformations
• In radiology, randomize kernel sharpness and slice thickness to mimic different reconstruction algorithms
• The model learns features invariant to these perturbations, effectively building a domain-agnostic representation
• Often combined with Test Time Augmentation (TTA) for additional robustness

A lightweight, inference-time adaptation method that aligns the feature statistics of target domain batches to match source domain distributions. Techniques include:

• Adaptive Batch Normalization (AdaBN): Re-estimate BatchNorm running means and variances on unlabeled target data before inference
• Instance Normalization: Normalize each image independently to remove instance-specific contrast and style information, which often carries domain-specific artifacts
• Deep CORAL: Aligns the second-order statistics (covariance matrices) of source and target feature distributions
• Advantage: no gradient updates required, making it suitable for edge deployment where retraining is impractical

A paradigm where the model continues to learn on each individual test sample at inference time using a self-supervised auxiliary task. For a detection model, this might involve:

• Rotating the test image and training the model to predict the rotation angle, forcing it to understand the image's structure
• Reconstructing masked patches of the test image to adapt feature extractors to its specific texture distribution
• Key benefit: adapts to each scan individually, handling inter-patient variability as well as inter-scanner variability
• Trade-off: adds computational latency at inference, which must be managed for real-time clinical workflows