List of Papers Browse by Subject Areas Author List
Abstract
Template-based segmentation, a widely used paradigm in medical imaging, propagates anatomical labels via deformable registration from a labeled atlas to a target image, and is often used to compute volumetric biomarkers for downstream decision-making. While conformal prediction (CP) provides finite-sample valid intervals for scalar metrics, existing segmentation-based uncertainty quantification (UQ) approaches either rely on learned model features, often unavailable in classic template-based pipelines, or treat the registration process as a black box, resulting in overly conservative intervals when applied directly in output space. We introduce ConVOLT, a CP framework that achieves efficient volumetric UQ by conditioning calibration on properties of the estimated deformation field from template-based segmentation. ConVOLT calibrates a learned volumetric scaling factor from deformation space features. We evaluate ConVOLT on template-based segmentation tasks involving global, regional, and label volumetry across multiple datasets and registration methods. ConVOLT achieves target coverage while producing substantially tighter intervals than output-space conformal baselines. Our work paves way to exploit the registration process for efficient UQ in medical imaging pipelines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2641_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/matthewyccheung/convolt
Link to the Dataset(s)
Learn2Reg Grand Challenge: https://learn2reg.grand-challenge.org/Datasets/
BibTex
@InProceedings{CheMat_Efficient_MICCAI2026,
author = { Cheung, Matt Y. AND Veeraraghavan, Ashok AND Balakrishnan, Guha},
title = { { Efficient Conformal Volumetry for Template-Based Segmentation } },
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
Structure volumes from an individual’s MRI were once estimated by warping the scan to an atlas so that volumes can be estimated by summing the Jacobians of the deformation over the labels in the atlas. This work uses a ridge regression with various features derived from the deformations to predict how to rescale these volume estimates so they are more accurate. Confidence intervals are estimated from a separate calibration set.
- 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 proposed method mostly gives tighter confidence intervals on test data than the other methods investigated.
- 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 manuscript presents a simple idea in a difficult way, with a information missing about how the ridge regression was done in practice. How were hyper-parameter(s) derived?
More recent approaches don’t try to segment structures using pure image registration. There is usually some form of machine learning involved as well, which will limit the usefulness of the proposed approach. There is a large literature on multi-atlas label propagation for refining segmentations, and there are other hybrid methods for combining deep earning segmentation with the outputs from nonlinear registration.
- 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 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 problem setup should say which way the warping is done: atlas to individual or individual to atlas?
As few people at MICCAI will be familiar with the concepts, I think the introduction should say what conformal prediction actually does (rather than use impenetrable text like “finite-sample valid coverage guarantees for scalar metrics”).
The use of $x$ to denote a spatial location (as in $J_i(x)$), as well as a set of features (as in Eq. 1) makes the manuscript harder to follow than it needed to be.
Does $X_{\mathrm{test}}$ refer to the same thing as $X_{n+1}$?
Don’t display brain images pointing face down (Figure 1). Is the “deformation field” shown in Figure 1 a map of Jacobians?
- 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 proposed method has limited applicability given that relatively few people use purely registration-based methods for structure segmentation.
- Reviewer confidence
Somewhat confident (2)
- [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
Review #2
- Please describe the contribution of the paper
Improves uncertainty quantification of volumetric measurements from template-based segmentation by predicting the multiplicative regression coefficient from deformation features.
- 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 novel, although based on known techniques.
- 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 approach is demonstrated only on volumetric biomarkers. Some aspects of the method are mentioned only in the experimental part and without much detail, namely the regressor, the similarity function and the features used for the regression.
The effect of the different features should be studied more, beyond just showing the graph in Figure 3 with hard to interpret labels.
- 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
In References, citing preprints should be avoided if possible.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The work is reasonably novel and experimentally validated.
- Reviewer confidence
Somewhat confident (2)
- [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.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes a conformal prediction (CP)-based uncertainty quantification (UQ) method for template-based segmentation, termed ConVOLT. The method aims to address two limitations in existing CP-based UQ approaches: their reliance on learned model features and the tendency to produce wide intervals when CP is applied directly in the output space. ConVOLT conditions CP on statistics derived from the deformation field by training a regressor that predicts the scaling factor between ground-truth volumes and predicted volumes. CP is then applied using the absolute error of this scaling factor regressor as the nonconformity score, yielding prediction intervals for the predicted volumes. The authors evaluate ConVOLT on three datasets using two types of registration methods and compare it against three CP baselines. The results show that ConVOLT achieves the desired coverage while producing narrower prediction intervals on most datasets.
- 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.Novelty: The proposed solution is well suited to the application domain. The authors incorporate deformation field statistics into the conformal prediction (CP) framework, addressing a gap that, to my knowledge, has not been explored in existing methods. The manner of incorporation is also meaningful rather than superficial. In particular, the choice of nonconformity score, based on the absolute error of the volume scaling factor, is domain-specific and differs from standard residual- or quantile-based approaches. This design is supported by a clear rationale, namely that multiplicative errors in the deformation propagate multiplicatively to the scalar target. 2.Rigorous Analysis: The authors provide a comprehensive evaluation by analyzing both coverage and interval size across global, regional, and label-wise volumetry. In addition, they examine the regression model to provide interpretability, with results that support the usefulness of deformation field statistics in datasets where ConVOLT shows improved performance. The paper also offers a reasonable explanation for the comparatively weaker performance on the NLST dataset when using the Demons registration method.
- 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.Formulation: The proposed method relies on an intermediate regressor to predict the volume scaling factor from deformation field statistics, and then applies conformal prediction to the residuals of this regressor. It is not clear why this two-step approach is necessary. An alternative formulation could directly condition the conformal prediction procedure on the deformation-derived features. For example, by using quantile regression with these features to predict intervals for the volumes. This would eliminate the need for the intermediate regression step. The authors could clarify why the current design provides advantages over a more direct conditioning approach. 2.Data Split: The data split (10/15/5) for ThoraxCBCT raises concerns as it results in a test set of only 5 samples. This makes it difficult to draw reliable conclusions about coverage and interval efficiency. With 5 samples, coverage can only be evaluated in increments of 20%, and this is reflected in a relatively large standard deviation of coverage (0.14), indicating high variability. Additionally, the calibration sets is unusually large relative to the dataset size (0.4/0.4/0.2, 10/15/5), raising questions about why such a large calibration set is needed and whether the method requires it to achieve stable coverage. 3.Baselines: The authors mention CQR (conditional quantile regression) as a baseline but cite Conformalized Quantile Regression. Can the authors clarify whether Conformalized Quantile Regression was actually used? Additionally, it is unclear what features were used for CQR — the only indication provided is that “residual volumetric variability is largely explained by volume magnitude itself.” 4.Ablation: Which features are considered localized and which are global? Additionally, can the authors explain the reduction in prediction interval size for VoxelMorph on ThoraxCBCT?
- 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
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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The major factors that led to my overall score are the novelty and domain relevance of the proposed method, as well as the rigorous evaluation provided. The paper makes a meaningful contribution by incorporating deformation field statistics into the conformal prediction framework, addressing a gap in existing methods. The choice of nonconformity score, based on the volume scaling factor, is well-justified for the application domain. The authors provide thorough analyses across global, regional, and label-wise volumetry, and they examine model interpretability, which strengthens the evidence for their approach.
While some aspects could be clarified. For instance, the necessity of the intermediate regressor, the small test set for ThoraxCBCT, and details of the CQR baseline. These do not outweigh the contribution of introducing a domain-specific conformal prediction method that demonstrates improved coverage and narrower prediction intervals.
- Reviewer confidence
Somewhat confident (2)
- [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 my comments on the important aspects of the paper (Formulation, Baselines, Features Used); therefore, I stand by my previous decision to accept it.
Author Feedback
We sincerely thank the reviewers and AC for the constructive feedback. We are glad that reviewers found our work novel and well evaluated (R2, R3). We focus on the top concerns.
Applicability/timeliness (R1,M): We respectfully clarify that registration/template-based volumetry remains relevant despite modern direct segmentation. ConVOLT targets settings where registration remains central: atlas-defined anatomical consistency and atlas/template deployment (OASIS), and longitudinal/respiratory analyses requiring dense correspondence or regional volume change (ThoraxCBCT, NLST). Such settings remain actively studied in volumetric deformable registration (e.g., VoxelMorph/TransMorph/HyperMorph/SynthMorph, Learn2Reg). Moreover, ConVOLT is not restricted to classical “non-ML” registration: we evaluate using VoxelMorph, a learning-based registration method. R1’s point about multi-atlas and hybrid deep-registration methods is therefore complementary rather than disqualifying — any pipeline that estimates deformation fields exposes the geometric quantities that ConVOLT leverages for efficient volumetric UQ. For direct segmenters, an analogous extension would use logits, embeddings, feature maps, or probabilities.
Sample size for ThoraxCBCT (R3,M): We agree ThoraxCBCT is small. With 5 test cases, split-level coverage is coarse and variable; under a simple binomial approximation with marginal coverage near 0.9, test noise alone gives sqrt(0.9*0.1/5)=0.13, close to the observed SD=0.14.We included it because it is a clinically relevant low-sample registration benchmark from Learn2Reg. The 15-case calibration split was chosen because split CP at alpha=0.1 estimates an upper conformal quantile; fewer calibration cases would make this more discrete/conservative. This applies to all split-CP methods, not only to ConVOLT.
ConVOLT features (R1-R3,M): We clarify that ConVOLT features are listed in the “Experimental Protocol” with notation from the problem setup. ConVOLT features were chosen a priori to summarize quantities expected to reflect registration difficulty: Jacobian statistics for expansion/contraction, Jacobian quantiles and folding fractions for extreme local volume changes, displacement and grad log-J/divergence/curl summaries for deformation magnitude and spatial irregularity, and fixed-warped similarity residuals for residual mismatch. Global experiments compute features over the full mask; regional/label experiments compute the same feature types restricted to the region/boundary.
ConVOLT versus CQR (R3,M): We agree that CQR conditioned on deformation features is a reasonable alternative, but it lacks ConVOLT’s structural constraints. Mapping a conformal interval in beta-space back to volume preserves non-negativity and scales interval width with the predicted volume. It separates point correction from uncertainty calibration: one ridge regressor estimates an interpretable multiplicative correction, while split CP supplies the finite-sample-valid radius. It also reduces the estimation burden in small-data and label-wise settings by using a single regressor rather than two conditional quantile regressors per target. Finally, coefficients directly indicate deformation features associated with multiplicative volumetric bias, whereas CQR has separate lower- and upper-quantile models whose effects on interval size are less transparent.
Baselines (R3,M): SCP uses no conditioning features; LCP conditions on |Yhat^0| using local calibration; CQR is conformalized quantile regression and conditions on baseline volume/change. These output-space baselines intentionally do not use registration internals, because the central comparison is whether deformation-aware calibration improves over black-box/output-space calibration.
Minor clarifications (R1,R2): We used closed-form ridge regression with an unpenalized intercept and fixed lambda=1e-2.We used NCC similarity, a standard VoxelMorph-style objective.
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 goal of this work is to provide an atlas-based segmentation approach combined with conformal prediction to predict the volumes of structures of interest. Features from the deformation field between the atlas image with segmentation and the image to be segmented are used to regress the volumes of interest and the conformal approach provides confidence intervals.
Major strengths of the paper:
- Reasonable regression approach to predict volumes and their confidence interval for an atlas-based segmentation approach.
- Evaluation shower tighter confidence intervals than competing approaches.
Major weaknesses of the paper:
- While the approach might be relevant for atlas/template-based segmentation, modern segmentation approaches have lately been dominated by deep-learning segmentation approaches (as by and large direct segmentation is empirically easier than accurate registration). This raises concerns about the timeliness of / need for the proposed approach.
- Features that are being used for regression have not been clearly described; also with respect to the competing approaches.
- There are some concerns regarding sample sizes for the experiments.
Suggestions for rebuttal:
- Discuss the timeliness / need of the approach in context of modern deep-learning based segmentation approaches. In what settings would one still want to use this approach? Could it be used (in a modified form) for deep learning segmentation methods?
- Address sample size concerns.
- Clarify what features were used and how they were chosen (for your own method as well as for competing approaches).
- What is the rationale in performing regressions exclusively based on deformation features?
- Address remaining reviewer concerns as space permits.
- 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 goal of this work is to provide an atlas-based segmentation approach combined with conformal prediction to predict the volumes of structures of interest. Features from the deformation field between the atlas image with segmentation and the image to be segmented are used to regress the volumes of interest and the conformal approach provides confidence intervals.
Major strengths of the paper:
- Reasonable regression approach to predict volumes and their confidence interval for an atlas-based segmentation approach.
- Evaluation shows tighter confidence intervals than competing approaches.
Initial concerns for this paper revolved around the following points: 1.While the approach might be relevant for atlas/template-based segmentation, modern segmentation approaches have lately been dominated by deep-learning segmentation approaches (as by and large direct segmentation is empirically easier than accurate registration). This raises concerns about the timeliness of / need for the proposed approach. 2.Features that are being used for regression have not been clearly described; also with respect to the competing approaches. 3.There are some concerns regarding sample sizes for the experiments.
The authors have addresses these concerns satisfactorily in the rebuttal. Specifically,
1.Template-based volumetry is the target for the proposed approach. 2.Features are provided in the experimental protocol section (and presumably could be better highlighted in a final version). 3.Sample sizes for ThoraxCBCT have been justified based on the desired to include the results “because [they are] a clinically relevant low-sample registration benchmark from Learn2Reg.
The reviewers who assessed the work after the rebuttal also recommended 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.
Accept
- Please justify your recommendation.
After the rebuttal, this paper received overall positive recommendations. Reviewer 3 explicitly stated that all three previously raised concerns had been addressed, and Reviewer 2 maintained an Accept recommendation. Reviewer 1 did not resubmit an updated review. After revisiting the paper in light of the reviews and rebuttal, I find that the proposed method offers a novel contribution in the context of conformal prediction. Although the limited size of the ThoraxCBCT dataset remains a weakness, it does not outweigh the overall merit of the work.
In light of the post-rebuttal discussion and the overall reviewer consensus, I recommend Accept. For the camera-ready version, the authors are encouraged to further improve the clarity of the presentation and to incorporate the reviewers’ constructive suggestions where appropriate.
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.
The paper is interesting, particularly because it moves beyond output-space conformal prediction and instead explores learned model features as more appropriate signals for uncertainty estimation. This direction is important for medical imaging, where directly applying CP in the output space can often lead to overly wide and less useful prediction intervals. The reviewers raised some questions regarding the timing and positioning of the work, but the paper appears timely and relevant to the medical imaging community. The empirical results are also strong and support the effectiveness of the proposed approach. Overall, this is a solid contribution, and I recommend acceptance.