Source-linked AI summary
CHILD: Human-in-the-Loop OOD Detection for Safe Clinical Deployment
Jinlun Ye, Kaiyue Lu, Runhe Lai, Xinhua Lu, Jia-Xin Zhuang, Ruixuan Wang
TL;DR
Clinical OOD detection must handle unseen disease categories that can produce overconfident, incorrect predictions, while automatic test-time adaptation raises safety concerns. CHILD selectively uses sparse physician feedback and retrieval-based score calibration for streaming detection; across four medical benchmarks, it reduces FPR95 by over 12% on average when combined with a state-of-the-art detector.
Problem
Unseen disease categories can produce overconfident, incorrect predictions, and automatically adapting OOD detectors raises safety concerns in clinical deployment.
Source: §1 · pp. 1–2Method
CHILD selects high-risk samples for limited physician review and calibrates OOD scores using compact caches of reviewed ID/OOD features without parameter updates.
Source: §1 · p. 2Results
Across four medical benchmarks, CHILD reduces FPR95 by over 12% on average when combined with a state-of-the-art detector.
Source: §1 · p. 2Takeaways & Limitations
The results support CHILD as a practical step toward safer medical AI systems with human oversight.
Source: §4 · p. 9In experiments, benchmark annotations serve as a controlled proxy for physician-confirmed feedback.
Source: §2.1 · p. 3
Abstract
from arXiv · showhide
Out-of-distribution (OOD) detection is critical for safe deployment of medical AI systems. Recently, test-time adaptation (TTA) has emerged as a new paradigm for OOD detection, automatically adjusting detector behavior during deployment. However, such automatic adaptation mechanisms may raise safety concerns in safety-critical clinical environments. While physician oversight can mitigate these risks, it is resource-intensive and must be judiciously allocated. To reconcile safety with efficiency, we propose CHILD, a training-free framework designed to enhance streaming OOD detection via sparse human feedback. Operating under strict budget constraints, CHILD employs an adaptive risk-aware sample selection mechanism to pinpoint only the most decision-uncertain samples for review. Crucially, it maximizes the utility of this sparse feedback through a retrieval-based score calibration module, which refines model predictions using a compact feature cache without any parameter updates. Extensive experiments on four medical benchmarks demonstrate that CHILD turns limited supervision into significant reliability gains: with a sparse feedback budget of only 5%, it reduces the average FPR95 from 72.63% to 60.26% and improves AUROC from 75.53% to 81.85%, consistently outperforming state-of-the-art baselines. Our code is publicly available at https://github.com/figec/CHILD.
1 Introduction
Clinical OOD detection must address unseen disease categories that can produce overconfident, incorrect predictions, while automatic test-time adaptation raises safety concerns. CHILD uses selective physician feedback and score calibration for streaming OOD detection under a limited feedback budget.
1 Introduction
- Unseen disease categories can elicit overconfident, incorrect predictions, making OOD detection important for safe medical AI deployment.Source: §1 · pp. 1–2
- TTA adjusts detection behavior from unlabeled test data, but clinical use raises safety concerns, while prior human-feedback work assumes offline access to the full test set.Clinical data arrive sequentially, so decisions must be made online using past observations.Source: §1 · p. 2
- CHILD selects high-risk, decision-uncertain samples for limited physician review and calibrates OOD scores using compact caches of reviewed ID/OOD features without parameter updates.Source: §1 · p. 2
- Across four medical benchmarks, CHILD reduces FPR95 by over 12% on average when combined with a state-of-the-art detector.Source: §1 · p. 2
- The framework targets streaming OOD detection with strict feedback budgets and combines selective queries with lightweight score-level calibration.Source: §1 · p. 2
2 Method
CHILD frames streaming OOD detection with a strict physician-feedback budget and combines risk-aware review selection with score-level calibration using reviewed samples, without parameter updates.
2.1 Preliminaries and Problem Setup
- Test samples arrive sequentially as an unlabeled ID/OOD stream, and physicians may review at most ρ% under a strict budget, with feedback mapped to model-defined classes.The system has no access to the complete test set in advance.Source: §2.1 · p. 3
2.2 Overview
- CHILD operates on an arbitrary base OOD detector, interleaving selection of decision-uncertain samples for limited physician review with calibration using the resulting feedback.The two components operate jointly as each sample arrives.Source: §2.2 · p. 3
2.3 Risk-Aware Sample Selection
- Otsu’s threshold on accumulated scores tracks a streaming reference boundary; a quantile-based width defines an adaptive interval that expands toward the more ID-like side.This targets decision-uncertain samples associated with more consequential OOD-as-ID errors.Source: §2.3 · pp. 4–5
2.4 Score Calibration
- Similarity to cached physician-confirmed ID/OOD features supplies a soft score adjustment, while confidence-gated hard assignment combines with it without parameter updates.The hard-assignment gate uses a semantic confidence threshold; the scaling factor aligns calibration terms with the base score.Source: §2.4 · pp. 5–6, §2.3 · p. 5
3 Experiment
Experiments evaluate CHILD on four medical benchmarks under a 5% feedback budget and compare its performance across detectors, components, feedback ratios, and hyperparameters. Results report gains with OODD and robustness across feedback budgets and parameter ranges.
3.1 Experiment Setup
- The evaluation uses SD-198, ISIC-4, NCT-CRC, and BreakHis, with a 5% feedback budget and causal streaming calibration.Feedback from a sample is used only to calibrate subsequent samples; results are averaged across the NCT-CRC and BreakHis splits.Source: §3.1 · pp. 6–7
3.2 Result Analysis
- On four benchmarks, CHILD achieves the best AUROC and FPR95 when combined with OODD.Table 1 reports percentage results, with lower FPR95 and higher AUROC preferred.Source: §3.2 · p. 7, Table 1 · p. 7
- Across other OOD detectors, CHILD’s results further support its general applicability.Figure 2 presents average performance with CHILD integrated into different detectors.Source: §3.2 · p. 7, Figure 2 · p. 8
- Each ablation component contributes consistent gains, and enabling all components gives the best results; the risk-interval gain is less pronounced on NCT-CRC.On NCT-CRC, samples selected within the risk interval can have similar characteristics, limiting boundary-focused querying; Table 2 distinguishes risk-aware from random selection.Source: §3.2 · p. 7, Table 2 · p. 8
- Right-side interval selection is stronger and more stable across feedback budgets, while AUROC approaches saturation around 1% feedback and gains beyond 5% are marginal.The comparison considers selecting samples to the left of the reference boundary, to its right, or symmetrically on both sides.Source: §3.2 · p. 8, Figure 3 · p. 8
- On ISIC-4, CHILD outperforms the best baseline across broad ranges of β, Qα, and η, indicating low sensitivity to these hyperparameters.The tested ranges are β = [0.01, 1.0], Qα = [0.3%, 4%], and η = [0.65, 0.95]; Figure 4 identifies OODD as the best baseline.Source: Figure 4 · p. 8, §3.2 · p. 9
4 Conclusion
CHILD combines sparse physician feedback with training-free score calibration for medical OOD detection. Experiments across multiple medical benchmarks report improved reliability under limited feedback.
- CHILD is a training-free human-in-the-loop framework for OOD detection motivated by clinical deployment.Source: §4 · p. 9
- Under a strict feedback budget, CHILD selectively queries high-risk samples and incorporates sparse physician feedback through lightweight score-level calibration.Source: §4 · p. 9
- Experiments on multiple medical benchmarks report consistent OOD detection improvements under limited feedback, supporting CHILD as a practical step toward safer medical AI with human oversight.Source: §4 · p. 9