Test-Time Adaptation

The most common and lightweight form of TTA. Instead of updating model weights, this method re-estimates the running mean and variance statistics in Batch Normalization layers using the current target batch.

• Mechanism: The model switches from accumulated source statistics to target-adaptive statistics.
• Benefit: Extremely fast; no backpropagation required.
• Use Case: Adapting to scanner-specific intensity shifts in CT or MRI volumes.
• Limitation: Less effective for severe distribution shifts that require feature-level adaptation.

A fully test-time optimization objective that updates model parameters by minimizing the entropy of the model's predictions on unlabeled target data.

• Assumption: The model's decision boundary should pass through low-density regions of the target distribution.
• Process: The model makes a prediction, computes the entropy of the softmax output, and takes a gradient step to make the prediction more confident.
• Risk: Can lead to model collapse if the target batch is not diverse, causing the model to always predict a single class.
• Variants: Often combined with a diversity regularizer or class-balancing prior to prevent collapse.

The model generates hard pseudo-labels for the unlabeled target data and then retrains itself using these artificial labels as supervision.

• Iterative Process: Predict → Filter high-confidence predictions → Retrain → Repeat.
• Filtering Criterion: Only pseudo-labels exceeding a confidence threshold (e.g., 0.9) are used for the update.
• Advantage: Directly optimizes the task loss (e.g., cross-entropy) rather than a proxy like entropy.
• Medical Imaging Relevance: Effective for adapting a tumor segmentation model to a new hospital's imaging protocol where the anatomy is consistent but the texture differs.

A dual-task framework where the model is trained on both a primary task (e.g., classification) and a self-supervised auxiliary task (e.g., rotation prediction) during source training. At test time, the model updates its shared feature extractor using only the auxiliary task on the target sample.

• Key Insight: The auxiliary task provides a supervised signal without needing labels.
• Architecture: Requires a shared backbone with two heads—one for the main task, one for the self-supervised task.
• Example: A model pre-trained to classify pathology and simultaneously predict image rotation can adapt to a new scanner by continuing to predict rotation on the target scans.
• Constraint: Requires modifying the original training procedure; not a plug-and-play solution for any pre-trained model.

Instead of updating the entire model, this method maintains a set of class prototypes (representative feature vectors) and adapts them to the target domain using the incoming feature distribution.

• Mechanism: Compute the mean feature vector for each class from source data. At test time, update these prototypes using a momentum scheme with target features assigned via nearest-neighbor or soft clustering.
• Classification: Target samples are classified by their cosine similarity to the adapted prototypes.
• Strength: Robust to label shift and avoids catastrophic forgetting of source knowledge.
• Application: Highly effective in medical imaging where disease class distributions vary across hospitals.

The choice of normalization layer fundamentally impacts TTA strategy. Vision Transformers (ViTs) use Layer Normalization, which computes statistics per-sample rather than per-batch, making standard BatchNorm recalibration inapplicable.

• BatchNorm: Statistics are batch-dependent; TTA can simply replace running estimates with target batch statistics.
• LayerNorm: Statistics are instance-dependent; TTA requires updating affine parameters or using prompt-based adaptation.
• Modern Approach: For ViTs, TTA often involves learning a lightweight visual prompt or adapter on the target data rather than modifying normalization.
• Takeaway: The TTA method must be matched to the backbone architecture.

A transfer learning technique that mitigates performance degradation caused by domain shift between a labeled source domain and an unlabeled or sparsely labeled target domain. Unlike TTA, domain adaptation typically requires access to target data during a dedicated training phase.

• Unsupervised DA: Aligns features without target labels
• Supervised DA: Uses a small set of labeled target samples
• Often employs adversarial training or moment matching

The statistical mismatch between the data distributions of the source domain used for training and the target domain encountered during deployment. In medical imaging, this is frequently caused by different scanner vendors, acquisition protocols, or patient populations.

• Covariate Shift: Change in input distribution P(X)
• Label Shift: Change in output distribution P(Y)
• Concept Drift: Change in the relationship P(Y|X)
• TTA directly addresses this without retraining

The challenge of training a model on one or several source domains that can robustly generalize to entirely unseen target domains without any additional adaptation. Unlike TTA, domain generalization does not access target data at all.

• Uses data augmentation to simulate domain shifts
• Employs meta-learning across multiple source domains
• Goal: Learn domain-invariant features
• Complementary to TTA; often combined for maximum robustness

A lightweight TTA method that updates the running mean and variance statistics of a model's Batch Normalization layers using target domain data. This corrects for covariate shift without modifying learned weights.

• Only adapts normalization statistics, not parameters
• Requires minimal compute and no backpropagation
• Effective when shift is primarily in image contrast or intensity
• Often the first TTA baseline tested in medical imaging pipelines

The task of identifying input samples at inference time that differ substantially from the model's training distribution. Critical for flagging unseen pathologies, scanner artifacts, or corrupted images before a diagnostic prediction is made.

• Uses softmax confidence scores or energy-based models
• Prevents silent failures on anomalous inputs
• Often paired with TTA to first adapt, then detect
• Essential for clinical safety and regulatory compliance

Techniques designed to standardize the appearance of medical images acquired from different scanner vendors or acquisition protocols to create a unified domain for model training. Unlike TTA, harmonization is typically applied as a pre-processing step.

• Histogram matching aligns intensity distributions
• CycleGAN performs unpaired image-to-image translation
• ComBat harmonization removes scanner-specific batch effects
• Reduces the magnitude of domain shift TTA must later correct