List of Papers Browse by Subject Areas Author List
Abstract
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.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1242_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/figec/CHILD
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YeJin_CHILD_MICCAI2026,
author = { Ye, Jinlun AND Lu, Kaiyue AND Lai, Runhe AND Lu, Xinhua AND Zhuang, Jia-Xin AND Wang, Ruixuan},
title = { { CHILD: Human-in-the-Loop OOD Detection for Safe Clinical Deployment } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes an alternative to test-time adaptation (TTA)-based OOD detection methods by incorporating clinical expert feedback into the OOD detection pipeline. This approach aims to classify clinical samples for which the statistical OOD detector is uncertain, particularly those that lie near the boundary between in-distribution and out-of-distribution data. The paper supports its methodological contribution with experiments demonstrating a significant improvement in OOD detection performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper proposes a somewhat novel method addressing an important aspect of OOD detection in safety-critical medical imaging, where near-boundary samples are difficult to detect. The novelty lies in how these samples are selected and adapted for human evaluation. 2.The paper presents sufficient experimental results, supported by strong ablation studies, thorough evaluation, and fine-tuning.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.There is less clarity regarding latency improvements when comparing offline TTA methods, such as OODD [26] with the proposed online method, in which clinicians handle only boundary samples as a stream of images arrives. 2.Another important question is whether the human-in-the-loop component has been evaluated properly. An ablation study involving multiple human evaluators to assess the reliability and consistency of their decisions would also be critical. 3.The numbers in Table 1 don’t report the variance after using different random seeds during test time.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The Fig. 4(b) has a typo in its horizontal axis caption, where it should be Q_{\alpha}.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a solid methodological contribution supported by sufficient experimental results. However, some open questions remain, and it would be helpful if the authors could address them during the rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The author’s rebuttal was sufficiently able to answer my questions. Therefore, I would recommend acceptance.
Review #2
- Please describe the contribution of the paper
The paper proposes CHILD (Clinical Human-In-the-Loop Decision), a training-free framework designed to improve streaming Out-of-Distribution (OOD) detection under strict human-feedback budget constraints. It achieves this through two main mechanisms: an adaptive risk-aware sample selection module that identifies decision-uncertain samples for physician review, and a retrieval-based score calibration module that adjusts model predictions using a feature cache of these human-reviewed samples without parameter updates.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The authors accurately identify a critical flaw in current Test-Time Adaptation (TTA) methods: automatic, unsupervised adjustments can lead to unpredictable behavior and progressive error accumulation in safety-critical domains. Leveraging naturally occurring physician oversight as a controlled feedback mechanism is a highly pragmatic approach to AI safety. 2.The use of Otsu thresholding dynamically applied to the empirical score distribution to locate the ID/OOD decision boundary is both mathematically sound and computationally lightweight. 3.The framework demonstrates substantial improvements when applied to strong baseline detectors. For example, when paired with OODD, it significantly lowers the FPR95 and increases AUROC across diverse datasets like Skin-40, ISIC-4, NCT-CRC, and BreakHis.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The text states that the ID and OOD feedback caches grow incrementally as new feedback is obtained. Because the calibration step relies on computing the maximum cosine similarity across the entire cache, the inference latency will scale linearly ($O(N)$) with the number of reviewed samples. The paper completely ignores cache management (e.g., FIFO queues, core-set selection) which is a critical algorithmic requirement for any streaming system intended for long-term deployment. 2.Equation 9 introduces a calibrated score $S_{final} = s_t + \beta(\delta_t + H_t)$. The baseline score $s_t$ and the similarity difference $\delta_t$ are not guaranteed to be in the same bounded range, especially across different base detectors (e.g., Energy vs. LogitGap). Fixing $\beta=0.08$ seems highly empirically fitted to the specific datasets tested and raises concerns about generalization without hyperparameter tuning on new target domains. 3.Streaming data evaluation is notoriously sensitive to the arrival order of the data. The paper presents average FPR95 and AUROC scores, but entirely lacks standard deviations, confidence intervals, or performance bounds over multiple random permutations of the data streams. 4.The framework relies on a 5% budget of labeled data. To prove the efficacy of the streaming calibration, the authors must compare CHILD against an offline baseline where the base detector is simply fine-tuned or calibrated using a randomly selected 5% subset of the test data prior to deployment. Without this, it is unclear if the performance gain is due to the sequential risk-aware querying or simply the injection of 5% ground-truth data.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The algorithmic approach relies on max cosine similarity in Equation 6.This nearest-neighbor approach is highly susceptible to feature space outliers. Have you considered a $k$-NN averaging approach to ensure that a single noisy human annotation does not permanently corrupt the hard-assignment gate? Additionally, releasing the code pipeline will be critical for verifying the streaming data loaders.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper addresses a highly important issue in clinical AI deployment: the safe adaptation of OOD detectors using limited human feedback. However, the score is capped at a Reject because the experimental design lacks rigorous statistical validation for streaming data (e.g., variance across different stream orderings). Furthermore, the lack of an algorithmic strategy to manage the linearly growing memory cache undermines the claim that this is ready for “real-world clinical deployment”.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The main contribution of the paper is a proposed online HITL OOD detection framework CHILD
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
The method seems to be quite practical. The design is training-free and can wrap around any base OOD detector
-
The proposed method is quite easy to understand, and therefore may be more practical for clinicians
-
The evaluation tests on four medical benchmarking datasets with extensive baselines, showing good results
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
The system is quite simple, and the technical contribution may not fully surpass that expected from a MICCAI paper
-
I would like some clarification regarding the human feedback. “ID/OOD” is defined with respects to the model training distribution, whereas a clinician typically observes the medical context of the image, not the model’s training support. In practice, it seems more realistic for physicians to provide to provide a diagnosis, category or “within intended scope” judgement, which is then mapped to ID/OOD by the study protocol. Please can the authors clarify the annotation protocol and justify why direct physician identification of ID vs OOD is feasible in real deployment?
-
the HITL setup may not be fully validated with real humans. I do not see a real human annotator study or a workflow showing how clinicians would actually provide this feedback in practice.
-
the ID/OOD setup for the experimental evaluation of categorical splits (i.e., some classes ID and rest OOD) is reasonable for benchmarking but may be too tidy for trying to represent a real deployment of such a system. I think real OOD shifts more often involve site/scanner/protocol changes, image artifacts, etc
-
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I choose weak reject due to my concerns on the technical sophistication may not reaching the bar of MICCAI, and my concerns regarding the human feedback
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers for their insightful feedback. R1Q1: The compared TTA baselines, including OODD, are online methods, not offline methods. We did not claim latency improvement over OODD. CHILD is applied on top of OODD to provide safer, controlled score-level calibration with sparse expert feedback. Profiling shows OODD+CHILD adds only 0.03 ms/sample, i.e., 0.36% overhead over OODD. R1Q2: To proxy inter-rater inconsistency, we randomly flip 10%/20% of queried feedback labels. CHILD still achieves 79.82/78.57 AUROC, compared with 81.42 under clean feedback and 75.37 for OODD, showing robustness to imperfect evaluator feedback. Real multi-evaluator validation will be discussed as future clinical work. R1Q3: We repeated streaming evaluation with 3 random seeds, each producing a different test-stream permutation. OODD+CHILD achieves 60.72±0.41 FPR95 / 81.42±0.40 AUROC, compared with OODD’s 72.26±1.05 / 75.37±0.62.We will add mean±std results for all methods in Table 1.R2Q1: Max-cosine calibration has linear cost in cache size. In our sparse-feedback setting, the cache stores only reviewed samples; the maximum cache sizes are 330/137/215/16 on the four benchmarks. For long-term scalability, we added FIFO bounded-cache analysis. With total cache cap K=50, cost becomes O(K). On the largest benchmark, the capped cache achieves 73.65/78.68 FPR95/AUROC, retaining most of the gain of the unbounded cache 68.91/81.55 and still clearly better than OODD 85.8/71.61. R2Q2: \(\beta\) is detector-specific, not dataset-specific. The role of \(\beta\) is to align the bounded feedback-based correction with the score scale of the base detector. Sensitivity analysis shows CHILD remains better than OODD over a broad range, with FPR95/AUROC ranging from 76.34/76.80 to 59.88/83.52.On an additional artifact/acquisition-shift setting, the same OODD-specific \(\beta\) improves OODD from 80.45/70.83 to 69.97/78.30, indicating the gain is not due to dataset-specific tuning. R2Q3: Same as R1Q3.R2Q4: Our ablation in Table 2 already shows the effectiveness of risk-aware querying. Following the reviewer’s suggestion, we further added an Offline-Random-5% baseline, evaluated on the same remaining 95% stream. It obtains 70.04/76.94 FPR95/AUROC, while CHILD achieves 60.82/81.66, confirming that the improvement mainly comes from risk-aware selection rather than simply injecting 5% GT labels. R2Q5: We tested k-NN averaging. It achieves 59.74/81.95 FPR95/AUROC, slightly better than max cosine 60.72/81.42, suggesting improved robustness to isolated outliers/noisy cache entries. Since it introduces an extra hyperparameter k, we keep max cosine as the main design. As stated in the abstract, full code will be released. R3Q1: CHILD is simple but effective and novel. Its contribution lies not in architectural complexity, but in formulating budgeted streaming human-in-the-loop OOD detection and designing a causal mechanism that allocates sparse expert feedback to high-risk boundary samples and uses it for training-free score calibration. R3Q2: Clinicians provide category feedback, mapped to ID/OOD by the predefined ID class set. In experiments, benchmark GT labels serve as a controlled proxy for such feedback, with ID/OOD mapping defined by the benchmark split. R3Q3: Our work focuses on the algorithmic question: how sparse feedback is allocated and used. In practice, clinicians provide category feedback, which is mapped to ID/OOD and stored in the feedback cache for future calibration. Real multi-clinician validation is an important next step. The feedback-noise analysis in R1Q2 shows CHILD remains effective under imperfect evaluator feedback. R3Q4: We added a preliminary domain/acquisition-shift experiment using Gaussian noise and JPEG compression as two representative non-categorical shifts. OODD obtains 80.45 FPR95 / 70.83 AUROC, while OODD+CHILD improves to 69.97 / 78.30.This indicates that CHILD is not limited to tidy categorical splits.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Please prioritize addressing the key issues, including scalability concerns such as cache growth and inference latency, clarifying the realism and validation of the human-in-the-loop setup (annotation protocol and feasibility of ID/OOD feedback), and evaluating robustness in the streaming setting (e.g., order sensitivity, variance, and comparison to an offline 5% labeled baseline).
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper proposes CHILD (Clinical Human-In-the-Loop Decision), a training-free framework for improving streaming OOD detection under strict human-feedback budget constraints. The approach is innovative, practical, and addresses an important deployment-relevant problem, as also acknowledged by the reviewers. The authors’ rebuttal further clarifies several concerns raised during the review process. One remaining issue is the computational cost of computing the maximum over an increasingly large cache at each inference step, which is an O(N) operation and could affect efficient deployment. However, the authors’ response is reasonable: the max-cosine calibration cost is linear in cache size, and in the proposed sparse-feedback setting, the cache contains only reviewed samples, with maximum cache sizes of 330/137/215/16 across the four benchmarks. This makes the cost practically acceptable in the evaluated setting. This limitations appear to be reasonable directions for future work rather than fatal flaws. Overall, the paper is solid and is recommended for acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
Two reviewers recommended rejection, while one recommended acceptance. The primary concerns raised by the reviewers centered on the insufficient experimental evaluation and analysis. Although the authors reported additional results during the rebuttal process, reporting new experimental results at this stage is not permitted under the MICCAI review policy. Furthermore, the rebuttal did not lead to a consensus among the reviewers in favor of acceptance. Based on the remaining concerns regarding the evaluation, the recommendation is to reject the paper.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Despite some remaining problems int he experimental setup, I agree with the reviewers that the paper presents a practical way to address the need for data annotation when using TTA in clinical settings, and I believe it could ignite meaningful and practical future work.