List of Papers Browse by Subject Areas Author List
Abstract
Medical image segmentation models commonly binarize voxelwise scores using a fixed threshold, such as 0.5. However, this heuristic provides no finite-sample guarantee against clinically consequential under-segmentation, particularly in 3D lesion segmentation, where lesions are sparse relative to the background. We propose Conformal Lesion Segmentation (CLS), a model-agnostic post-processing framework for controlling voxel-level false-negative risk. Given a user-specified FNR tolerance ε and violation probability α, CLS uses a held-out calibration set to compute, for each case, the smallest mask-expansion parameter for which the voxel-level FNR does not exceed ε. A finite-sample-corrected conformal quantile of these critical values is then used to determine a global binarization threshold. Under exchangeability, CLS guarantees that a new test case satisfies the prescribed FNR tolerance with probability at least 1 − α. Because the critical parameter is chosen as the smallest feasible expansion, CLS also limits the associated increase in false positives, although FPR is not itself formally controlled. We evaluate CLS on six 3D-LS benchmarks across five backbone models, demonstrating its superior statistical validity and predictive performance, and providing potential guidance for deploying risk-aware segmentation models in real-world clinical applications. Code can be found at https://github.com/binyutan/CLS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2825_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/binyutan/CLS
Link to the Dataset(s)
KiTS21 dataset: https://kits-challenge.org/kits21/
LiTS dataset: https://competitions.codalab.org/competitions/17094
NIH-LN ABD dataset: https://www.cancerimagingarchive.net/collection/ct-lymph-nodes/
LIDC-IDRI dataset: https://www.cancerimagingarchive.net/collection/lidc-idri/
MDSC-Colon dataset: http://medicaldecathlon.com/
MDSC-Pancreas dataset: http://medicaldecathlon.com/
BibTex
@InProceedings{TanBin_Conformal_MICCAI2026,
author = { Tan, Binyu AND Wang, Zhiyuan AND Duan, Jinhao AND Xu, Kaidi AND Shen, Heng Tao AND Shen, Fumin AND Shi, Xiaoshuang},
title = { { Conformal 3D Lesion Segmentation with Balanced Risk Control } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The manuscript introduces “conformal lesion segmentation” (CLS), a split conformal prediction (SCP) method which selects a binary segmentation threshold in order to minimize the false positive rate (FPR) within a specified false negative rate (FNR) constraint. Overall, this method produces lower FNRs than using a heuristic threshold, without compromising the FPR.
- 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 claim that their methodology is the first to adapt SCP to binary segmentation settings. 2.The topic of conformal prediction for segmentation is timely and would be of great interest to the uncertainty quantification (UQ) community within MICCAI. 3.The proposed methodology is applied post model development, which is important for clinical settings as UQ is unfortunately often an afterthought. 4.The methodology provides a FNR guarantee, which is especially important in tasks like lesion segmentation. Importantly, empirical results demonstrated that the guarantee did not substantially deteriorate the FPR. 5.The manuscript is well written and formulated: (1) large assortment of experiments: 6 datasets, 5 models; (2) claims a release of code; (3) appropriately grounds the methodology in a clinical context in the introduction; (4) contains a figure that effectively functions as a graphical abstract; (5) multiple figures demonstrate the variability of the results, instead of just rely on point estimates of performance; and (6) an ablation study is performed on calibration dataset size.
- 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 manuscript would benefit from a stronger connection to the existing literature on conformal prediction for medical image segmentation. For example, the introduction does not adequately situate the work within recent developments. 2.The proposed prediction compactness (PC) metric is difficult to interpret and appears to be largely redundant with standard measures such as FPR or precision. Under a fixed FNR constraint, PC effectively reduces to a normalized measure of predicted mask size, which does not provide additional insight beyond existing metrics and may obscure clinically relevant aspects of segmentation quality. 3.The vast majority of the results were only presented for one dataset. For example, the results of the comparison with heuristic thresholding, the evaluation of how the FPR was affected, the PC results, and the ablation study on calibration dataset size were only presented for KiTS21.While I understand the space is limited, this makes the reader question whether the results extend to new modalities/segmentation tasks. 4.Lack of comparison with other threshold-selection methods. The work basically demonstrates that a SCP-selected decision threshold improves FNR while not completely compromising FPR over a heuristic threshold of 0.5 on the KiTS dataset. Comparing against other threshold-selection methodologies, or discussing why comparison is not feasible, would strengthen the manuscript. 5.The methodology is mostly evaluated in terms of FNR/FPRs. As changing the decision threshold changes the predictive segmentation mask, it would have been informative to also use metrics that measure segmentation performance (such as DSC, HD, etc. )
- Please rate the clarity and organization of this paper
Good
- 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
1.Abstract: “(i. e. , 0.5)” not “(i. g. , 0.5)”. Please clarify what the fixed threshold is used for (i. e. , decision making). 2.Figure 2: Please add more explanation for subfigure (b), such as what CMR means. In addition, please clarify what annotations the text is referring to for P-FNR. 3.Before equation 6, please explain the purpose behind the chosen quantile. 4.Figures 3 and 4 contain very small and condense subplots. 5.Why was epsilon raised to 0.4 in Figure 5, when it was 0.2 elsewhere? Changing epsilon without justification makes the reader wonder if the results for epsilon=0.2 were not as good. 6.In the “Robustness to Calibration-Test Split Ratios” section, please add a caveat that the interquartile range was larger for smaller proportions, especially for models such as SamMed3D.
- 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?
Adapting SCP for medical image segmentation holds promise and will be of interest to the UQ community at MICCAI.
While I would not recommend acceptance for a journal due to the lack of comparison to other threshold selection methods and the lack of presenting results on 4 out of 5 datasets, promising MICCAI works could be limited to proofs of concept.
In a rebuttal, I would like to see (1) bringing the work into the context of related conformal prediction in medical imaging segmentation literature, (2) a more tempered presentation of the merits of PC, and (3) the additional comments fixed.
- 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.
Adapting SCP for medical image segmentation to control the FNR is novel and will be of interest to the uncertainty quantification community at MICCAI.
The main drawbacks to this paper of a lack of comparison to other threshold selection methods (outside of a 0.5 heuristic) and the lack of presenting results on 4 out of 5 datasets remain.
Review #2
- Please describe the contribution of the paper
The paper introduces Conformal Lesion Segmentation (CLS), a post-processing framework that extends split conformal prediction to 3D medical image segmentation. CLS adaptively calibrates decision thresholds to guarantee that the false negative rate (FNR) on test data does not exceed a user-specified tolerance ε with probability at least 1-α, while minimizing false positive rate (FPR). The method addresses a critical gap in current segmentation models that rely on fixed heuristic thresholds (typically 0.5), which often lead to uncontrolled FNR—particularly problematic in 3D lesion segmentation where lesions are sparse and complete misses have serious clinical consequences.
- 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.
Novel methodological contribution: CLS is the first work to provide statistical guarantees for FNR control in segmentation through conformal prediction. The formulation of FNR-specific and FPR-specific loss functions as nonconformity scores is technically sound and innovative. The derivation of the critical threshold ti as the solution to a constrained optimization problem (minimizing FPR while bounding FNR) elegantly adapts conformal prediction theory to the segmentation domain.
Strong theoretical foundation: Theorem 1 provides rigorous statistical guarantees for FNR control, with a clear proof establishing that the calibrated threshold ensures Pr(L_FNR_test(ˆt) ≤ ε) ≥ 1−α. The exchangeability assumption is appropriate and standard for conformal prediction methods. Comprehensive experimental validation: The evaluation spans six diverse 3D medical segmentation datasets (KiTS21, LiTS, NIH-LN ABD, LIDC-IDRI, MDSC-Colon, MDSC-Pancreas) and five state-of-the-art backbone architectures (Med3D, nnUNet, UNETR, Swin-UNETR, SAM-Med3D). This breadth demonstrates generalizability across anatomical regions and model designs.
Clinically motivated problem formulation: The paper effectively identifies and articulates a critical real-world problem—uncontrolled FNR in lesion segmentation leading to missed diagnoses. The distinction between voxel-level FNR (V-FNR) and presence-level FNR (P-FNR) is particularly relevant clinically, as complete lesion misses (P-FNR) are especially dangerous.
Novel evaluation metric: The introduction of Prediction Compactness (PC) as a metric for spatial efficiency under risk constraints provides a practical tool for comparing uncertainty-aware segmentation models. This addresses an important gap in how to evaluate segmentation quality when operating under formal risk guarantees.
Robustness analysis: The experiments examining different calibration-test split ratios (Figure 5b) demonstrate that CLS maintains statistical validity even with limited calibration data (10%), which is crucial for practical deployment scenarios.
- 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.
Limited comparison with relevant baselines: While the paper compares CLS against fixed thresholding (t=0.5), it lacks comparison with other uncertainty calibration or threshold adaptation methods commonly used in medical image segmentation. Specifically, there is no comparison with: Temperature scaling or Platt scaling for calibration Other risk-control methods such as those based on Bayesian neural networks or ensemble uncertainty Recent conformal prediction variants for segmentation (e.g., [1] Angelopoulos et al. on image classification uncertainty sets, or segmentation-specific conformal methods if they exist)
Insufficient ablation studies: The paper does not investigate: The effect of different ε values on the same dataset with fixed α The interaction between model calibration quality and CLS performance—poorly calibrated models may require larger calibration sets Computational overhead of the calibration procedure compared to standard inference
Missing clinical utility validation: While the paper claims clinical relevance, there is no validation with clinical readers or assessment of how the predicted compactness (PC) or risk levels (α) correlate with actual clinical decision-making. The clinical significance of the FNR reductions shown in Table 1 is asserted but not demonstrated through clinical outcomes or reader studies.
Assumption limitations not fully discussed: The exchangeability assumption may be violated in real clinical deployment due to: Domain shift between calibration and test data (different scanners, protocols, patient populations) Temporal drift in data distribution The paper briefly acknowledges this but does not provide empirical validation of robustness to such violations, which is critical for safety-critical clinical applications.
Incomplete discussion of failure modes: Figure 3 shows occasional ECR violations below the theoretical bound 1−α. The paper attributes this to “finite-sample variability,” but does not quantify: The frequency or severity of these violations Whether certain datasets or models are more prone to violations Practical consequences when the guarantee fails
Limited exploration of the ε-α trade-off space: The experiments primarily use ε=0.2, α=0.2 or ε=0.4, α=0.2.A more systematic exploration of how these hyperparameters interact and their practical selection guidance would strengthen the work.
- 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?
None
- 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
Typos and formatting issues: Abstract, introduction: “i.g.” should be “e.g.”? Page 7, Table 1: “clibration” should be “calibration” in Figure 4 caption
Methodological clarification needed: Specifically, clarify why taking the quantile of t_i values provides the desired coverage guarantee.
Experimental details: The exact train/calibration/test splits are not specified—what proportion of the original training data is used for D_train vs D_cal?
- 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?
The paper addresses an important and underexplored problem in medical image segmentation—statistical control of false negative rates—through a theoretically grounded conformal prediction framework.
However, the work is limited by insufficient comparison with alternative calibration and risk-control methods, a lack of clinical validation, and unaddressed practical concerns regarding distribution shift and failure modes, as well as some minor typo errors.
In all, I recommend Weak Reject.
- 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 authors have adequately addressed the reviewers’ concerns by leveraging existing content and have outlined plans to further supplement and improve the work
Review #3
- Please describe the contribution of the paper
This paper proposes a post-hoc conformal calibration framework, Conformal Lesion Segmentation (CLS), for 3D lesion segmentation. Starting from a pretrained segmentation model, the method uses a held-out calibration set to compute per-case critical thresholds based on an FNR-specific loss, then selects a global test-time threshold via a conformal quantile. The stated goal is to guarantee, under exchangeability, that the voxel-level false negative rate remains below a user-specified tolerance with probability at least 1–alpha, while keeping the false positive rate low. The paper evaluates CLS on six public 3D lesion segmentation benchmarks and five backbone models, showing improved FNR control relative to standard fixed thresholding. It also proposes “prediction compactness” as an auxiliary risk-aware benchmarking measure.
- 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) Clinically meaningful objective. The paper focuses on controlling false negatives in lesion segmentation, which is an important safety-related failure mode in medical imaging. This is a better-motivated target than relying on a default threshold such as 0.5, especially in sparse-lesion settings where missed lesions can be clinically consequential.
(2) Simple and practical post-hoc formulation. The method is conceptually simple: given a pretrained segmentation network, it calibrates a decision threshold using held-out data, without requiring retraining. This makes it potentially easy to deploy on top of existing systems.
(3) Clear theoretical motivation. The paper gives a concise conformal-style argument under exchangeability and derives a threshold from calibration-case critical values. The theoretical framing is easy to follow and aligned with the broader conformal prediction literature.
(4) Broad empirical coverage. The evaluation spans six 3D lesion segmentation datasets and five architectures, which is stronger than a single-dataset proof of concept and suggests the method is not backbone-specific.
(5) Useful calibration perspective. The paper highlights that standard thresholding can lead to very poor FNR behavior in lesion segmentation and that post-hoc risk calibration can materially change that trade-off. This is a worthwhile perspective for safety-aware segmentation deployment.
- 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 novelty is not fully convincing relative to prior conformal segmentation work. The paper claims to be, to the best of the authors’ knowledge, the first post-processing approach to provide FNR guarantees in segmentation, but this seems too strong. Prior work has already adapted conformal prediction to image segmentation and explicitly addressed the false-negative / coverage side of the problem through prediction sets, including binary segmentation settings and biomedical segmentation settings, e.g., [1–3].
(2) The experimental baseline is too weak for a methodological paper. Most of the empirical comparison is against standard heuristic thresholding (for example, 0.5), which is useful as a sanity check but not sufficient to establish SoTA novelty. Given the existing literature on conformal segmentation, the paper should compare directly against representative prior conformal approaches rather than only against fixed-threshold inference, e.g., [1–3].
(3) The paper’s contribution appears closer to a task-specific adaptation of existing conformal ideas than to a fundamentally new methodological framework. The main idea is to calibrate a segmentation threshold on a held-out set so that false negative risk is controlled with high probability. That is a sensible and practical idea, but the manuscript does not clearly articulate why this adaptation is substantially different from prior segmentation-oriented conformal formulations based on score thresholding, prediction sets, or inner/outer confidence regions, e.g., [1–3].
(4) The empirical evaluation is reasonably broad in terms of datasets and backbones, but it is narrow in terms of metrics and comparisons. The paper focuses mainly on ECR, voxel-level FNR/FPR, and prediction compactness, while standard segmentation metrics such as Dice, lesion-wise sensitivity, or boundary-aware metrics are not emphasized enough. For a MICCAI audience, it is important to understand whether the proposed risk control comes at an acceptable cost in overall segmentation quality.
(5) The proposed “prediction compactness” metric is not yet fully justified as a meaningful new benchmark. It is intuitively related to prediction size and false-positive behavior, but the paper does not clearly establish why it should be preferred over, or interpreted differently from, more standard notions of mask size, over-segmentation, or set size that are already discussed in prior conformal segmentation literature, e.g., [1–3].
(6) Some claims are stronger than the evidence presented. For example, the paper states that the method consistently enforces the desired control, but the manuscript also notes that empirical compliance can occasionally fall slightly below the nominal lower bound in practice. This does not invalidate the method, but the wording should be more careful and the practical limitations should be discussed more explicitly.
(7) The related work section is underdeveloped for a methodological submission. In particular, the paper would benefit from a more careful discussion of prior segmentation-specific conformal methods and from a clearer explanation of what is genuinely new here: the risk definition, the threshold construction, the target guarantee, or the empirical application. At present, the positioning overstates novelty and understates prior SoTA, e.g., [1–3].
Refs
[1] Mossina, L., and Friedrich, C. “Conformal Semantic Image Segmentation: Post-hoc Quantification of Predictive Uncertainty.” In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPR Workshops), 2024. [2] Mossina, L., and Friedrich, C. “Conformal Prediction for Image Segmentation Using Morphological Prediction Sets.” In Medical Image Computing and Computer Assisted Intervention (MICCAI), 2025. [3] Davenport, S. “Conformal Confidence Sets for Biomedical Image Segmentation.” arXiv preprint arXiv:2410.03406, 2024.
- 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
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 found the paper interesting and potentially useful, especially because it targets false negative control in lesion segmentation, which is clinically meaningful. The proposed post-hoc calibration is simple, practical, and evaluated across multiple datasets and backbones. However, I am not convinced the paper demonstrates sufficient methodological novelty over prior conformal segmentation work. Existing papers already study conformal prediction for segmentation using thresholded score sets, coverage-oriented control of missed regions, and morphology-based conformal prediction sets for segmentation. In that context, the present method appears closer to a task-specific adaptation and empirical study than to a clearly distinct methodological advance.
A second major concern is the experimental framing: the method is compared mainly to heuristic thresholding rather than to prior conformal segmentation baselines. For a methodological MICCAI submission, this makes it difficult to judge the real increment over the state of the art.
Because of these two points, I place the paper below the acceptance threshold. While the problem is important and the paper has some practical value, the core concerns relate to novelty and evaluation, and these are substantial rather than marginal.
- 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 the reviewers for their thoughtful comments. We are encouraged by their recognition of the clinical importance of controlling false negatives, the relevance of post-hoc calibration, and the empirical breadth across datasets and backbones. We will revise the manuscript to better clarify the positioning, guarantee, and scope of CLS.
Clarifying the contribution. To R1/R2/R3.We do not claim that CLS is the first use of conformal prediction for segmentation. Its specific contribution is to address a clinically motivated risk objective not directly targeted by standard thresholding or many existing conformal segmentation formulations: controlling the probability that the per-case voxel-level FNR exceeds a user-specified tolerance. CLS calibrates a single post-processing threshold so that the deployed output remains a standard binary mask while satisfying a finite-sample marginal conformal guarantee under exchangeability. This is distinct from methods that construct voxel-wise prediction sets, uncertainty regions, or morphology-based conformal masks. We will revise the abstract, introduction, and contribution list accordingly.
Relation to heuristic and validation-based thresholding. To R1/R2/R3.We include the fixed-0.5 comparison because it is a common operational choice that directly reveals the key clinical failure mode: sparse lesions may receive low predicted probabilities, causing high V-FNR and complete misses. CLS does not merely lower the threshold; it replaces an uncontrolled heuristic with a statistically calibrated rule. A validation-tuned threshold may reduce empirical FNR on a held-out set but does not provide the same finite-sample marginal guarantee for unseen exchangeable cases. We will discuss this distinction alongside validation-based thresholding and recent conformal segmentation methods in the related work.
Clarifying the guarantee. To R1/R2.The theorem provides a marginal per-test-case guarantee: for an exchangeable test example, the probability that its voxel-level FNR-specific loss exceeds the tolerance is at most \alpha. It is not a deterministic guarantee that every finite test split must show empirical compliance above 1-\alpha. Split-level fluctuations below the nominal 1-\alpha line in Fig. 3 therefore do not contradict the conformal statement; they reflect finite-sample variability of empirical compliance estimates. We will revise the theorem explanation, ECR discussion, and figure caption to eliminate this ambiguity.
Prediction compactness. To R1/R2/R3.PC is not intended to replace standard metrics such as Dice, sensitivity, HD95, or NSD. Its role is narrower: once FNR tolerance is controlled, PC summarizes the spatial expansion required to achieve that control, serving as an auxiliary footprint measure for comparing backbone conservatism under FNR-controlled calibration. We will reposition it as an auxiliary metric to be interpreted alongside V-FNR, P-FNR, and V-FPR.
Clinical scope and assumptions. To R1/R2.Our claim is methodological: CLS provides a post-hoc mechanism to control missed-lesion risk under the standard exchangeability assumption. We agree that reader studies and outcome-based validation are needed before clinical deployment. We will explicitly note that scanner/protocol shifts, temporal drift, and cross-institutional deployment may weaken exchangeability, and that future work should incorporate domain-shift detection or site-specific recalibration.
Finally, we will correct all noted presentation issues, including typos, fixed-threshold notation, the quantile explanation, P-FNR notation, calibration/test split description, and figure readability. Overall, CLS offers a practical, statistically grounded way to convert existing 3D lesion segmentation models into risk-aware systems that explicitly prioritize reducing missed lesions while preserving a standard binary-mask output.
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.
The topic of conformal prediction in segmentation problems is unanimously considered as important and timely by the reviewers. However there are some key issues repeatedly raised that would require proper justification / clarification at rebuttal stage and notably in terms of the contextualisation with respect to other conformal prediction methods for segmentation and the associated comparison design. Although some issues are major, I am keen to hear from the authors at rebuttal stage to see whether this is enough to allow acceptance
- 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.
In the rebuttal, the authors have convincingly addressed the positioning of the proposed method and clarified its main intention. While the evaluation may still be in early stages, the method and intention are interesting enough to warrant further discussion at the conference
Meta-review #2
- 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.
Following the rebuttal, R1 and R2 found the clarifications convincing, with R2 explicitly recommending acceptance. Although R3 maintained a Weak Reject (desiring comparisons against other conformal baselines), the overarching consensus is that the strong clinical motivation, solid theoretical grounding, and extensive empirical evaluation make this a valuable and reliable methodological contribution.
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.
2 reviewers concluded that the rebuttal was successful, I concur with them (one reviewer did not submit the postrebutal decision). Rebuttal period was successful, and novelty was emphasized better.