Domain Shift

The most common form of domain shift in medical imaging, where the input distribution P(X) changes but the conditional label distribution P(Y|X) remains the same.

• Scanner vendor differences: Siemens vs. GE MRI machines produce different texture patterns
• Acquisition protocol variation: Contrast-enhanced vs. non-contrast CT scans
• Patient population drift: Pediatric vs. geriatric anatomy distributions

This shift causes models to encounter feature values outside their training range, leading to miscalibrated predictions.

Occurs when the prior probability of classes P(Y) differs between source and target domains, while P(X|Y) remains stable.

• Prevalence mismatch: A model trained on a balanced dataset encounters real-world screening where disease prevalence is < 1%
• Severity distribution: Training on advanced-stage pathology but deploying on early-stage detection
• Comorbidity patterns: Different co-occurring condition frequencies across hospital systems

Label shift directly impacts calibration and requires prevalence-aware threshold adjustment.

The most severe form of domain shift where P(Y|X) itself changes—the relationship between features and labels differs across domains.

• Annotation protocol divergence: Different radiologists use varying criteria for lesion malignancy grading
• Evolving diagnostic standards: Updated BI-RADS or LI-RADS classification systems
• Population-specific pathology: Disease presentation varies across ethnic groups

Concept drift cannot be resolved by simple normalization and requires explicit domain adaptation or retraining.

Shift caused by the physical image formation process rather than biological variation.

• Slice thickness: 1mm vs. 5mm CT reconstructions alter partial volume effects
• Reconstruction kernel: Sharp vs. smooth kernels change texture and edge characteristics
• Dose variation: Low-dose protocols introduce quantum noise patterns absent in standard-dose training data
• Motion artifacts: Cardiac or respiratory motion patterns vary by patient compliance and scanner speed

These shifts are deterministic and can often be addressed through physics-based harmonization.

The target domain contains entirely novel classes or conditions not present in the source training data.

• Rare pathology emergence: Novel disease presentations absent from historical training sets
• Implant and device artifacts: Surgical hardware creates image features never seen during pre-training
• Anatomical variants: Congenital anomalies outside the training distribution

This requires out-of-distribution detection mechanisms to flag unknown inputs rather than forcing incorrect predictions.

Progressive degradation of model performance as the target data distribution evolves over time.

• Scanner fleet upgrades: Gradual replacement of older MRI systems with newer models
• Protocol optimization: Radiology departments periodically update imaging guidelines
• Population demographics: Shifting age distributions in the served patient population
• Seasonal variation: Certain pathologies show different presentation patterns across seasons

Continuous monitoring and periodic recalibration are essential to detect and mitigate temporal drift before clinical impact occurs.