OOD Detection for Medical Imaging Foundation Models: Singapore Hospital Guide

Foundation models for medical imaging—large pre-trained networks that generate rich embeddings for downstream tasks—are entering clinical workflows in Singapore hospitals. They promise to reduce annotation burden, improve rare disease detection, and enable rapid adaptation to new tasks. But they carry a silent risk: when input data shifts away from training distributions—different scanner vendors, patient populations, or acquisition protocols—performance degrades unpredictably, often without visible warning.

This post is for clinical AI teams, hospital CIOs, and medical imaging engineers in Singapore evaluating or deploying foundation model pipelines. We examine recent research on out-of-distribution (OOD) detection methods, explain why distribution shifts are endemic in Asia-Pacific healthcare settings, and provide a practical framework for building safety guardrails into imaging AI systems.

Key takeaways

  • Foundation models for medical imaging degrade silently under distribution shifts from scanner changes, population differences, or protocol variations—common in Singapore's multi-site, multi-vendor hospital environment.
  • New OOD detection methods like Mahalanobis Residual Scoring [1] exploit frozen foundation model embeddings to flag unreliable predictions without retraining, enabling post-hoc safety layers.
  • Digital pathology foundation models face unique integration challenges [2]: whole-slide image scale, stain variability, and workflow fragmentation require careful OOD monitoring at multiple pipeline stages.
  • Effective OOD detection in production requires calibrated thresholds, clinician-facing uncertainty displays, and fallback protocols—not just model accuracy metrics.
  • Singapore hospitals should pilot OOD detectors on retrospective multi-site data before deploying foundation models in clinical workflows, especially for high-stakes tasks like cancer staging or pulmonary embolism triage.

Why foundation models fail silently under distribution shift

Foundation models—whether self-supervised vision transformers, contrastive learning models, or large-scale pre-trained CNNs—learn representations from massive datasets. When fine-tuned on hospital-specific tasks, they often outperform models trained from scratch, especially in data-scarce settings [4].

But medical imaging data is heterogeneous. A chest X-ray foundation model trained on North American academic centers may encounter:

  • Scanner vendor differences: GE vs. Siemens CT protocols produce subtly different noise profiles and contrast curves.
  • Population shifts: Singapore's ethnic diversity (Chinese, Malay, Indian, expatriate populations) introduces anatomical and disease prevalence variations not represented in Western training sets.
  • Acquisition protocol drift: A hospital switches from 1.5T to 3T MRI, or updates reconstruction algorithms, changing image statistics.
  • Rare pathologies: A foundation model trained on common pneumonias encounters a novel infectious outbreak (a scenario Singapore hospitals know well).

Under these shifts, foundation model embeddings remain high-dimensional and seemingly informative, but downstream classifiers built on them produce overconfident, incorrect predictions. Standard accuracy metrics on held-out test sets don't capture this—because the test set was drawn from the same distribution as training data.

How Mahalanobis Residual Scoring detects unreliable predictions

A recent preprint [1] introduces MaRS (Mahalanobis Residual Scoring), a post-hoc OOD detector designed specifically for frozen foundation model embeddings. The method combines two signals:

  1. Mahalanobis distance: Measures how far a test embedding is from the distribution of in-distribution training embeddings in feature space. Embeddings from unfamiliar inputs (new scanner, rare pathology) tend to lie far from the training manifold.
  2. Reconstruction residual: Uses a lightweight autoencoder to reconstruct embeddings; high reconstruction error indicates the embedding structure differs from training patterns.

By fusing these scores, MaRS flags predictions likely to be unreliable without requiring model retraining or access to OOD examples during development. This is critical for Singapore hospitals, where:

  • Foundation models are often vendor-supplied or pre-trained externally, with frozen weights.
  • Retraining on local data is expensive and may not be contractually permitted.
  • New distribution shifts (scanner upgrades, population changes) emerge post-deployment.

MaRS and similar methods (kNN-based detectors, energy-based scoring) enable a safety layer architecture: the foundation model generates embeddings, a downstream classifier makes predictions, and an OOD detector flags cases for human review or fallback protocols.

Digital pathology: where OOD detection gets harder

Digital pathology presents unique challenges for foundation model deployment and OOD detection [2]. A systematic review published this month highlights integration barriers:

  • Whole-slide image scale: Gigapixel images require multiple instance learning (MIL) pipelines, where OOD detection must operate on both patch-level embeddings and slide-level aggregations.
  • Stain variability: Hematoxylin and eosin (H&E) staining protocols vary across labs and even across batches within a lab, creating continuous distribution drift.
  • Workflow fragmentation: Pathology AI must integrate with laboratory information systems (LIS), slide scanners, and reporting tools—each a potential source of data format or metadata shifts.
  • Regulatory uncertainty: Singapore's HSA AI-SaMD pathways are still maturing for digital pathology; OOD detection may become a regulatory expectation for high-risk applications like cancer diagnosis.

For pathology foundation models, we recommend OOD monitoring at three pipeline stages:

  1. Slide ingestion: Flag slides with unusual stain intensity, tissue folding, or scanning artifacts before processing.
  2. Patch embedding: Detect patches with rare tissue types or staining patterns not seen in training.
  3. Slide-level prediction: Flag cases where patch-level OOD scores aggregate to high slide-level uncertainty.

This multi-stage approach prevents unreliable predictions from propagating through the pipeline and reaching pathologists as false confidence.

Pulmonary embolism risk stratification: when foundation models meet structured data

A recent preprint on pulmonary embolism (PE) risk stratification [3] illustrates another OOD challenge: multimodal distribution shifts. PE risk guidelines combine CTPA imaging, patient medical records, and blood tests. Foundation models trained on CTPA alone may produce embeddings that ignore critical clinical context (e.g., patient age, comorbidities, troponin levels).

When blood test results are missing—common in routine practice—models trained on complete data encounter a missingness distribution shift. The study found that vascular graph representations from CTPA, while powerful, are insufficient without clinical context. OOD detectors must flag not just unusual images, but unusual combinations of image and clinical features.

For Singapore hospitals deploying multimodal imaging AI, this means:

  • OOD detection must operate on fused embeddings (image + EHR), not just image embeddings.
  • Missing data patterns should trigger uncertainty flags, not silent imputation.
  • Fallback protocols should route high-uncertainty cases to senior clinicians or multidisciplinary teams.

Why this matters in Singapore and Asia

Singapore's healthcare AI deployment context amplifies OOD risks:

  • Multi-site clusters: Public healthcare clusters (NUHS, SingHealth, NHG) operate multiple hospitals with heterogeneous equipment and patient populations. A foundation model trained at one site may encounter distribution shifts at another.
  • Vendor diversity: Singapore hospitals use imaging equipment from Siemens, GE, Philips, Canon, and others—often within the same institution. Scanner-specific distribution shifts are the norm, not the exception.
  • Regional deployment: Singapore AI vendors often deploy models across Southeast Asia, where population genetics, disease prevalence, and imaging infrastructure differ sharply from Singapore.
  • Regulatory scrutiny: HSA's AI-SaMD framework and upcoming post-market surveillance requirements will likely mandate OOD monitoring for high-risk devices. Hospitals that build OOD detection into pipelines now will have smoother regulatory pathways later.

Without OOD detection, foundation model deployments risk silent performance degradation: models continue to produce predictions, but accuracy drops unnoticed until a clinical incident triggers retrospective review. This is unacceptable in healthcare.

A practical OOD detection framework for imaging AI pipelines

Based on recent research and our experience with clinical AI services in Singapore hospitals, we recommend this five-step framework:

1. Establish in-distribution baselines

  • Collect embeddings from your foundation model on a representative in-distribution validation set (same scanners, protocols, population as intended deployment).
  • Compute distribution statistics: mean, covariance (for Mahalanobis distance), or fit a density model (Gaussian mixture, normalizing flow).
  • Store these as OOD detector parameters—they define what "normal" looks like.

2. Choose OOD scoring methods

No single OOD detector works for all imaging tasks. Pilot multiple methods:

  • Mahalanobis distance [1]: Fast, interpretable, works well for foundation model embeddings.
  • kNN-based scores: Compute distance to k-nearest training embeddings; simple and effective.
  • Reconstruction-based: Train a lightweight autoencoder on in-distribution embeddings; flag high reconstruction error.
  • Energy-based models: Use the logit output of your classifier as an energy score; OOD inputs tend to have lower energy.

Evaluate on retrospective OOD examples: scans from a different scanner, a different hospital site, or a different time period. Measure AUROC for OOD detection.

3. Calibrate thresholds for clinical workflows

OOD scores are continuous; you must choose a threshold to flag cases for review. This is a clinical workflow decision, not just a statistical one:

  • High sensitivity (low threshold): Flag many cases, reduce missed OOD inputs, but increase radiologist review burden.
  • High specificity (high threshold): Flag fewer cases, reduce review burden, but risk missing subtle distribution shifts.

Work with radiologists to define acceptable false-positive rates. For high-stakes tasks (cancer staging, PE triage), err toward high sensitivity.

4. Build fallback protocols

When the OOD detector flags a case, what happens?

  • Human review: Route flagged cases to senior radiologists or a multidisciplinary team.
  • Fallback model: Use a simpler, more robust model (e.g., a rule-based system or a model trained on diverse data).
  • Defer decision: Mark the case as "uncertain" in the radiology report, prompting additional imaging or clinical correlation.

Document these protocols in your AI governance framework and train clinical staff on them.

5. Monitor OOD rates over time

OOD detection is not a one-time validation step—it's a continuous monitoring signal. Track:

  • OOD rate: Percentage of cases flagged per week/month. A sudden increase indicates a distribution shift (scanner upgrade, protocol change, population shift).
  • OOD score distribution: Visualize score histograms over time; gradual drift may precede sudden failures.
  • Clinical outcomes: For flagged cases that were reviewed, track diagnostic accuracy and adverse events.

Integrate OOD monitoring into your clinical AI drift monitoring infrastructure.

What to do next

  • Audit your imaging AI pipeline: Identify where foundation models are used (pre-trained backbones, vendor-supplied embeddings, transfer learning). Map potential distribution shift sources (scanner vendors, sites, protocols).
  • Pilot OOD detection on retrospective data: Before deploying a foundation model, test OOD detectors on scans from a different site or scanner. Measure detection performance and calibrate thresholds with radiologists.
  • Build OOD monitoring into your MLOps stack: Instrument your inference pipeline to log embeddings, compute OOD scores, and alert when OOD rates exceed thresholds. Treat OOD detection as a first-class production metric, like latency or throughput.
  • Document fallback protocols: Write clinical SOPs for OOD-flagged cases. Train radiologists and technologists on how to interpret uncertainty flags.
  • Engage with regulators early: If you're developing an AI-SaMD, discuss OOD detection with HSA reviewers. Post-market surveillance plans should include OOD monitoring as a safety signal.

If you're evaluating foundation models for medical imaging in Singapore and need help designing OOD detection strategies, start a project with our team.

FAQ

What's the difference between OOD detection and model calibration?

Model calibration adjusts prediction confidence scores to match true probabilities (e.g., if a model predicts 80% confidence, it should be correct 80% of the time). OOD detection identifies inputs that are fundamentally different from training data, where even a well-calibrated model may fail. Both are important: calibration improves trust in predictions on in-distribution data, while OOD detection prevents overconfident predictions on unfamiliar data. For medical imaging, you need both—calibration for routine cases, OOD detection for edge cases.

Can I use OOD detection with vendor-supplied foundation models?

Yes, and you should. Vendor-supplied models (e.g., pre-trained imaging backbones from commercial AI platforms) are often black boxes with frozen weights. You can't retrain them on your local data, but you can extract embeddings and build OOD detectors on top. Methods like Mahalanobis distance [1] and kNN scoring work with frozen embeddings. This is one of the few safety interventions hospitals can implement without vendor cooperation.

How do I handle continuous distribution drift, like gradual scanner aging?

Continuous drift (scanner calibration drift, population aging, seasonal disease patterns) requires adaptive OOD detection. Instead of fixed in-distribution baselines, update your OOD detector parameters periodically (e.g., quarterly) using recent data. Monitor OOD rates over time; a gradual increase signals drift. When drift is detected, consider retraining your downstream classifier or recalibrating your OOD thresholds. This is part of broader post-deployment drift monitoring practices.

Should OOD detection be part of HSA AI-SaMD submissions?

HSA's current AI-SaMD guidance doesn't explicitly mandate OOD detection, but it requires post-market surveillance and performance monitoring. OOD detection is a natural fit for these requirements—it provides an early warning signal for performance degradation. If you're submitting a high-risk imaging AI device (e.g., cancer diagnosis, stroke triage), including OOD detection in your risk management plan strengthens your submission. We expect regulators to increasingly expect OOD monitoring as the field matures.

Sources

[1] MaRS: Robust Out-of-Distribution Detection via Mahalanobis Residual Scoring. arXiv preprint, June 2026. https://arxiv.org/abs/2606.22649v1

[2] Chen Y, Qin X, Chen S. Foundation Models and AI Agents in Digital Pathology Imaging: A Systematic Review of Integration into the Clinical Workflow and Implementation Challenges. Journal of Imaging Informatics in Medicine, June 2026. https://pubmed.ncbi.nlm.nih.gov/42298093/

[3] Pulmonary Embolism Risk Stratification from CTPA and Medical Records: Vascular Graphs Are Not All You Need. arXiv preprint, June 2026. https://arxiv.org/abs/2606.25956v1

[4] Re-mixing Embeddings for Patient Augmentation in Data Scarce Multiple Instance Learning. arXiv preprint, June 2026. https://arxiv.org/abs/2606.25770v1