Out-of-Distribution Detection

A set of statistical methods that equip a diagnostic AI model with the ability to estimate the confidence of its own predictions. Unlike OOD detection, which flags anomalous inputs, uncertainty quantification operates on in-distribution data to identify ambiguous cases.

• Epistemic uncertainty: Model uncertainty due to lack of knowledge, reducible with more data
• Aleatoric uncertainty: Inherent noise in the data, such as ambiguous lesion boundaries
• Monte Carlo Dropout: A practical technique that enables uncertainty estimation by sampling multiple forward passes with dropout enabled at inference time

Together with OOD detection, uncertainty quantification creates a dual safety net: the system knows both when it's looking at something unfamiliar and when it's unsure about something familiar.

The continuous monitoring process that identifies when a deployed model's statistical properties or predictive performance degrade over time due to changes in the input data distribution. While OOD detection operates at the individual inference level, drift detection operates at the population level.

• Data drift: The input distribution P(X) changes, e.g., a new scanner vendor's reconstruction algorithm alters pixel statistics
• Concept drift: The relationship P(Y|X) between inputs and labels changes, e.g., new diagnostic criteria for a disease emerge
• Kullback-Leibler divergence and Maximum Mean Discrepancy are common statistical tests used to detect drift in production

Drift detection triggers model retraining or rollback, while OOD detection provides real-time per-inference safety.

A training methodology designed to produce a model that maintains robust diagnostic performance when deployed on data from entirely new hospitals or scanner vendors not seen during training. This is the proactive counterpart to OOD detection's reactive safety mechanism.

• Multi-source domain alignment: Training on diverse data from multiple institutions to learn invariant features
• Data augmentation with style transfer: Simulating scanner-specific variations during training
• Meta-learning approaches: Training the model to quickly adapt to new domains with minimal examples

While OOD detection says 'I don't recognize this,' domain generalization aims to make the model recognize more in the first place. The two techniques are complementary layers in a defense-in-depth strategy.

Feature attribution and saliency mapping techniques that make diagnostic model decisions auditable and interpretable by human radiologists. When an OOD detector flags an input, explainability methods help clinicians understand why the model was uncertain.

• Grad-CAM: Generates heatmaps highlighting image regions most influential to a classification decision
• SHAP values: Quantifies the contribution of each pixel or feature to the model's output
• Counterfactual explanations: Shows how an input would need to change to alter the prediction

Explainability transforms OOD detection from a binary 'reject' signal into an actionable clinical insight, showing radiologists exactly which anatomical regions triggered the anomaly flag.

A critical preprocessing step for CT images that rescales pixel intensity values to a standardized physical unit of radiodensity, ensuring consistent model input regardless of the scanner's reconstruction parameters. Improper normalization is a leading cause of false OOD triggers.

• Water = 0 HU, Air = -1000 HU: The fixed reference points for the Hounsfield scale
• Windowing: Clipping intensity ranges to specific tissue windows (e.g., lung, bone, soft tissue) before model input
• Scanner-specific calibration: Different manufacturers apply proprietary reconstruction kernels that shift intensity distributions

Robust normalization reduces the likelihood that a legitimate CT scan from an unfamiliar scanner is incorrectly flagged as out-of-distribution, improving the specificity of OOD detection in multi-vendor hospital environments.

A hybrid architecture that intelligently distributes AI workloads between local edge devices and cloud infrastructure. OOD detection plays a key role in this orchestration by determining which inferences can be trusted locally and which require escalation.

• Tier 1 (Edge): Run standard inference with OOD check; if confident, return result immediately at the scanner
• Tier 2 (Cloud): If OOD is detected, offload the study to a more powerful cloud model or queue for human review
• Bandwidth optimization: Only anomalous studies consume costly cloud bandwidth, reducing operational expenses

This architecture ensures that OOD detection doesn't create workflow dead-ends. Instead, flagged cases are automatically routed to the appropriate escalation path, maintaining clinical continuity.