CURE-OOD: Benchmarking Out-of-Distribution Detection for Cancer Survival Prediction on CT Imaging

> Paper: CURE-OOD: Benchmarking Out-of-Distribution Detection for Survival Prediction > Authors: Wenjie Zhao, Jia Li, Mingrui Liu, Jing Wang, Yunhui Guo > arXiv: 2605.00350 | 2026-04-29

The Medical AI Dilemma: "The Model Stops Working at New Hospitals"

Consider this scenario: a patient asks an AI system how long they will survive or whether their cancer will recur. The model predicts based on CT imaging, but it was trained at Hospital A and deployed at Hospital B, where different CT scanners, different parameters, and different image quality create distribution shift. The predictions become unreliable.

The deeper problem is that the model doesn't know what it doesn't know — it produces confident predictions even on OOD samples, and patients may make decisions based on false information. What's needed is OOD detection: identifying out-of-distribution images and rejecting unreliable predictions to protect patient safety.

CURE-OOD: The First OOD Detection Benchmark for Survival Prediction

CURE-OOD's core idea is to establish the first benchmark for OOD detection in cancer survival prediction, evaluating model reliability under distribution shift.

Key technical components:

• OOD detection: identifying out-of-distribution samples — judging whether the model has seen similar data before, not simply classifying
• Survival prediction: time-to-event prediction rather than binary classification, which is more complex and makes OOD detection harder
• CT imaging: 3D, high-dimensional medical data with diverse distribution shifts (scanner differences, protocol differences, patient population differences)
• Benchmarking: systematic evaluation of multiple OOD detection methods across multiple shift types, revealing strengths and weaknesses

The analogy: traditional AI is like a doctor who only "gives answers," while a system with OOD detection is like a doctor who can say "I haven't seen anything like this — let's run further tests." More honest, more reliable.

Why OOD Detection Is Critical for Medical AI

Without OOD detection:

• False confidence: models are confident even on OOD samples; patients misplace trust and make wrong decisions
• Safety hazards: wrong survival predictions can affect treatment choices, with psychological impact and serious consequences
• Trust crisis: once errors occur, patients distrust AI and doctors avoid using it, hindering deployment

With OOD detection:

• Honesty: the model knows what it doesn't know and doesn't guess blindly
• Reliability: predictions only on in-distribution samples; OOD cases are deferred to humans, enabling human-AI collaboration
• Trust: patients understand AI's limitations and doctors know when to trust it — safer deployment

Takeaways

If you develop medical AI or OOD detection systems, ask yourself:

1. Does my model know what it doesn't know? 2. Have distribution shifts been fully considered? 3. Will the model refuse to predict when uncertain? 4. Does OOD detection affect patient safety?

In the spirit of Feynman: knowing what not to do is as important as knowing what to do. The most dangerous quality of an AI system is not knowing what it doesn't know. In medical prediction, refusing to answer is sometimes more valuable than answering wrongly.

CURE-OOD reminds us: in medical AI, reliability matters more than accuracy. When AI learns to honestly say "I don't know," it transforms from a prediction machine into a trustworthy medical partner. The best medical AI isn't the most accurate — it's the one that knows best when to say "I'm not sure."