List of Papers Browse by Subject Areas Author List
Abstract
Quantitative brain biometry from MRI is essential for evaluating brain growth in preterm-born neonates, yet manual measurement is time-consuming and suffers from inter-rater variability. Existing automation approaches require dense voxel-wise annotations that are costly to obtain and sensitive to domain shifts. We present MEASURE (Measurement Estimation via Anatomical Slice Understanding and REgression), the first annotation-efficient deep learning framework for Kidokoro-protocol neonatal brain biometry. Our two-stage approach mirrors the clinical workflow: a ViT-based multi-task slice selector identifies optimal measurement planes, followed by a multi-task regression network with measurement-specific spatial attention that jointly predicts all biometrics without voxel-level supervision. Experiments on 470 developing Human Connectome Project subjects demonstrate that explicit plane selection consistently outperforms end-to-end 3D volumetric baselines, and our proposed predictor outperforms multi-task regression approaches on simulated thick-slice data.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4530_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LeeJiy_MEASURE_MICCAI2026,
author = { Lee, Jiyang AND Bae, Woori AND Kim, Dabin AND Lee, Jong-Min AND Kim, Seh Hyun},
title = { { MEASURE: Multi-Task Slice Selection and Regression for Neonatal Brain Biometry } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a two-stage framework for predicting neonatal brain biometry measurements. In the first stage, a Vit model is used for slice selection, identifying the optimal measurement planes by selecting the most appropriate slice for each measurement. However, rather than formulating the problem as a binary classification task, it was defined as a three-class classification problem, distinguishing between the optimal slice, its neighbouring slices, and all remaining slices, under the assumption that neighbouring slices also represent near-optimal measurement planes. In the second stage, a regression network is employed to estimate the biometric measurements, where mechanisms such as spatially adaptive routing, measurement-specific attention pooling, and anatomical attention regularization are used to improve the prediction. Although fetal brain biometry estimation has been extensively studied in the literature, neonatal biometry estimation has received comparatively less attention. Therefore, this work addresses an important gap and has the potential for significant impact.
- 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.
Although the motivation and clinical importance of the proposed method are clear, and the methodology is well formulated through a two-stage approach, first detecting the optimal measurement slice and subsequently estimating the measurements on those slices (thus avoiding the more challenging task of directly estimating 2D measurements from 3D data), while also incorporating different components in the network to improve estimation accuracy, the main strength of the paper lies in its strong end-to-end evaluation. The authors present a comprehensive evaluation, including the required statistical analysis for biometric measurement assessment (Pearson correlation, MAE, ICC, Bland-Altman), intra-rater analysis, ablation studies on different parts of the model, and comparisons with state-of-the-art methods. Additionally, the method is also applied to the FeTA dataset to demonstrate its generalization performance.
- 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.
Since the Kidokoro scoring system for assessing brain abnormalities is highly mentioned throughout the manuscript, being used to present the clinical motivation of the work, a more explicit connection between the results of the paper and this scoring system should have been provided. The Kidokoro scoring system does not necessarily rely on the specific biometric measurements extracted in this work, while also considering additional MRI findings. In fact, the Kidokoro score relies on a combination of MRI findings, such as signal abnormalities, structural alterations, and developmental features, where some of them can be inferred from the biometric measurements. Therefore, the paper could be more explicit about how the retrieved measurements can be used, or how the proposed methodology could be extended, to support the computation of a Kidokoro score.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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 scheme in Figure 1, particularly part (b), could be improved by explicitly including the lightweight CNN that encodes inter-slice difference maps. As currently presented, it appears that the through-plane augmentation is applied to the features F extracted by the ResNet, rather than being applied to the input channels and subsequently fused with F.
- Additionally, Figure 1 would benefit from the inclusion of input/output nomenclature (e. g. , F, F’, F_out) to better guide the reader through the mathematical description of the method.
- Although the mathematical formulation is well presented, the inputs to the ResNet are not entirely clear. How are the inputs for each plane combined before being fed into the ResNet? Moreover, for each plane, is the input a six-channel representation composed of the central slice, neighbouring slices, and coordinate channels? How are these coordinate channels defined (e. g. , Cartesian, radial, etc. )? Please clarify.
- The authors state on page 7 that “the predictor consumes top-2 candidates. ” However, the methodology section seems to indicate that the predictor takes only the top-1 candidate as input, represented by a six-channel input (2.5D input plus coordinate channels). Is this correct? Please clarify.
- In Table 1, within the ablation study, does the “base” model correspond to the version without attention pooling, SAFA, and AAR? This could be stated explicitly in the paper.
- Although it is presented as an additional experiment, the evaluation on the FeTA dataset would benefit from further clarification, particularly regarding the use of “GA only. ” How was the pipeline adapted to operate using only gestational age information? Please clarify.
- 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 addresses a clinically relevant problem and proposes a well-structured framework. A key strength is the comprehensive end-to-end evaluation. Overall, the completeness of the manuscript (clinical relevance, well-designed framework, and strong validation) put the manuscript above the acceptance threshold from the reviewers’ perspective.
- 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.
Regarding the reviewer’s concerns related to the Kidokoro assessment, the authors clarified that MEASURE automates only the quantitative biometric component, which represents just one part of the Kidokoro framework, instead of leading the reader to assume that the proposed methodology automatically estimates or extracts the full score. This clarification makes the manuscript more accurate and fair in the way the contribution is presented.
In addition, the authors addressed several points that were previously unclear to the reviewer, even if some of the responses were relatively brief.
Overall, considering that the topic is relevant, the methodology is well formulated through a two-stage approach, and the authors present a comprehensive evaluation, the reviewer considers that this work is above the acceptance threshold.
Review #2
- Please describe the contribution of the paper
The paper presents MEASURE, a two-stage deep learning framework for neonatal brain biometry that mimics the clinical workflow. It first performs multi-task slice selection using a ViT-based model, then predicts multiple biometric measurements via a regression network with measurement-specific attention. The approach is annotation-efficient, avoiding dense voxel-level supervision, and demonstrates strong performance on the dHCP dataset with comprehensive evaluation.
- 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 paper addresses a clinically important problem with strong alignment to real-world measurement workflows and eliminates the need for dense annotations, improving practical applicability. The proposed two-stage architecture is well motivated and thoughtfully designed. The work is supported by comprehensive and rigorous experimental evaluation, including cross-validation, multi-metric analysis, ablations, and comparison against human variability, with clear and reproducible presentation. In addition, Fig. 2 demonstrates high interpretability, where attention maps reflect the underlying biometry of interest.
- 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 evaluation of thick-slice conditions relies on simulated data rather than real clinical acquisitions, limiting practical validation. The study depends on single-expert annotations without inter-rater analysis, and comparisons with 3D methods may be biased due to explicit slice supervision. Furthermore, interpretability claims are only qualitatively supported, cross-domain generalization evidence is limited, and some architectural components (e.g., SAFA, AAR) show inconsistent gains.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(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 well-motivated and technically sound approach with strong experimental validation and clear clinical relevance. Although there are limitations in real-world validation and annotation diversity, the work is solid and demonstrates meaningful improvements.
- Reviewer confidence
Very confident (4)
- [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 #3
- Please describe the contribution of the paper
This paper presents MEASURE, a two-stage deep learning framework for automated neonatal brain biometry. The method decomposes the task into (1) slice selection using a ViT-based multi-task classifier and (2) measurement prediction via a ResNet-based regression model with additional modules (SAFA and measurement-specific attention pooling). Stage 2 uses a ResNet-34 backbone augmented with the proposed SAFA module — which combines through-plane difference maps with spatially adaptive fine/coarse routing — and measurement-specific attention pooling with AAR. A key motivation is to avoid voxel-level annotations by relying only on slice indices and scalar measurement labels.
- 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 well-motivated and annotation-efficient design. The proposed framework requires only slice indices and scalar measurements as supervision, avoiding voxel-level annotations entirely. This is a practically meaningful advantage for neonatal brain MRI, where expert annotation is costly and time-consuming.
2.Rigorous and transparent evaluation protocol. The paper presents a strong and well-structured evaluation, including 5-fold cross-validation, multiple complementary metrics (ICC, Pearson r, MAE), and statistical testing with Wilcoxon signed-rank tests and Benjamini–Hochberg FDR correction. Notably, the authors explicitly mark ablation cases where variants outperform the full model but are not statistically significant (◦), demonstrating a commendable level of transparency that is often lacking in similar works.
3.Inclusion of clinically relevant evaluation (Bland–Altman analysis). The use of Bland–Altman analysis provides additional clinical context beyond standard ML metrics, which is often missing in similar work. While not novel, this analysis helps interpret model performance relative to intra-rater variability and improves the practical relevance of the evaluation.
- 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.Cross-domain experiment does not support — and in fact weakens — the generalization claim. The FeTA evaluation is presented as evidence of architectural transferability, but the results are not convincing. The image-only configuration fails to converge, and even when combined with GA, image features provide only limited and inconsistent improvements, with no gains for some structures (e.g., HV, LCC). While the authors attribute this to limited data, the results more strongly indicate sensitivity to domain shift and data scale. As such, this experiment reads less as supporting evidence and more as a demonstration of the method’s limitations, weakening the overall claim of generalizability. 2.Single-site, single-protocol dataset severely limits generalizability. All experiments are conducted on a single dataset (dHCP) acquired under controlled conditions. This does not reflect real clinical variability, where neonatal MRI often involves lower resolution, thick slices, and motion artifacts. 3.Limited empirical contribution of SAFA and AAR. Although the authors explicitly mark non-significant results (◦) and provide partial explanations (e.g., interpretability trade-offs for AAR), the empirical evidence supporting these modules remains weak. In particular, SAFA shows statistically significant improvement only in a small subset of metrics, while AAR does not yield significant gains. The provided explanations are largely qualitative and do not sufficiently justify their inclusion as key components of the method. 4.The fine/coarse routing mechanism relies on under-justified heuristic gating. The mixing weight 𝑟 is defined using a Sobel-derived edge complexity map, a fixed threshold 𝜏, and a learnable temperature, but the paper does not adequately explain how the fixed threshold is chosen or how sensitive performance is to this design. This makes the routing mechanism appear heuristic rather than well-validated, especially given the low contrast and noise characteristics of neonatal MRI. 5.The proposed “attention” mechanism is closer to spatial weighting than relational attention. The measurement-specific attention operates as a spatial weighting (i.e., weighted pooling) over feature maps, rather than modeling interactions between spatial locations (e.g., via query-key-value attention). While effective in practice, this limits its ability to capture global dependencies and reduces the conceptual novelty of the module.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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?
The primary factor behind my weak reject recommendation is that the experimental evidence does not sufficiently support the paper’s claims of generalizability and methodological contribution.
Most importantly, the cross-domain (FeTA) experiment is not convincing and in fact weakens the generalization claim. The image-only configuration fails to converge, and even with GA, image features provide only limited and inconsistent improvements, with no gains for some structures (e.g., HV, LCC). While the authors attribute this to limited data, the results more strongly suggest sensitivity to domain shift and data scale rather than architectural robustness.
This concern is further compounded by the fact that all main experiments are conducted on a single dataset with homogeneous acquisition conditions, which does not reflect real-world clinical variability. As such, the paper does not provide sufficient evidence that the method would generalize beyond the specific setting considered.
In addition, the empirical contribution of the proposed modules is limited. The ablation results show that SAFA yields statistically significant improvements only in a small subset of metrics, while AAR does not demonstrate significant gains. Although the authors provide qualitative explanations (e.g., interpretability trade-offs), these do not sufficiently justify presenting these components as key contributions.
Finally, some design choices appear heuristic and insufficiently validated. In particular, the fine/coarse routing relies on a Sobel-based complexity map and a fixed threshold without justification or sensitivity analysis, and the proposed “attention” mechanism functions primarily as spatial reweighting rather than modeling meaningful spatial relationships. These factors further limit the novelty and robustness of the approach.
Overall, while the problem and overall framework are well-motivated, the current evidence is not strong enough to fully support the paper’s broader claims, leading to a weak reject recommendation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The rebuttal addresses several concerns in a reasonable and professional manner, particularly by clarifying terminology, acknowledging dataset limitations, and reframing the FeTA experiment as a limited transfer study rather than definitive evidence of broad generalization. However, my overall assessment remains unchanged.
The main reason is that the empirical support for some of the paper’s key methodological components remains weak. In particular, AAR is presented throughout the paper as an important methodological contribution, yet the ablation results show no statistically significant performance gains and in several cases slightly worse results. In the rebuttal, the authors clarify that AAR is intended primarily as an interpretability-oriented regularizer rather than a component designed to improve predictive accuracy. While this clarification is helpful, the paper does not provide sufficient evidence that the resulting attention maps offer clinically meaningful interpretability benefits that justify the observed performance trade-off. As a result, the practical necessity and impact of AAR remain unclear.
Overall, I appreciate the authors’ thoughtful responses and believe the work is promising, but I still find the current evidence insufficient to fully support the paper’s broader methodological and clinical claims.
Author Feedback
##### We apprecia all reviewers for recognizing the overall framework, clinical relevance, comprehensive evaluation, and transparent reporting. ####
[R1] Kidokoro scoring connection. Full Kidokoro scoring also requires signal abnormalities, structural alterations, and qualitative findings beyond biometry. We will clarify that MEASURE automates the quantitative biometric component, is a module that can be integrated with quantitative detectors toward comprehensive automated scoring.
[R1] Clarifications. (a) Input: 6ch = 3 slices (2.5D) + 3 coordinate channels (x,y,radial). (b) Selector returns top-2; predictor randomly samples one as center slice. (c) Base = ResNet-34 + GAP + prediction heads. (d) GA-only is a univariate linear regression baseline per FeTA challenge protocol, not an input to our pipeline.
[R2] 3D comparison framing. This comparison tests clinical workflow-based decomposition benefits and plane-aware supervision, not intrinsic 2D-vs-3D superiority. We will revise wording.
[R2] Interpretability. Measurement-wise pooling maps directly determine predictions via measurement-specific weighted pooling, providing spatial explanations of which regions contribute to each biometric. We agree this is demonstrated qualitatively (Fig. 2) rather than via quantitative localization.
[R2] Single-expert annotations. Table 2 compares model–expert with intra-rater variability (>3-month re-measurement) via Bland–Altman LoA ratios. We acknowledge inter-rater analysis will strengthen the validation.
[R2&R3] Simulated thick-slice. Our simulation provides preliminary robustness evidence, not a substitute for real clinical MRI. We will make this limitation explicit.
[R2&R3] FeTA cross-domain. This was not intended as proof of broad generalizability, but as a transparent transfer test under substantial domain shift (term neonatal to younger fetal, N=470 to N=70). In context, most FeTA challenge methods did not outperform GA-only regression, our Image+GA improves MAE for 3/5 targets with clear gains for sBIP and TCD. We will clarify this as limited but non-trivial evidence of architectural adaptability.
[R2&R3] Role of SAFA and AAR. We will clarify the component hierarchy: measurement-specific pooling is the main performance driver (Table 1, largest ablation gap). SAFA provides selective, statistically supported benefits (FDR-corrected for BPW and IHD, p=0.036), consistent with its through-plane context role. AAR only works with the presence of measurement-specific pooling (Table 1, w/o pooling ablation is also excluding AAR), and is not intended as an independent component to improve regression accuracy, Rather, it regularizes pooling maps via symmetry/smoothness/distinctiveness, consistent with the structured behavior qualitatively illustrated in Fig. 2.We will revise framing accordingly.
[R3] Attention terminology. We agree that our module is not relational query–key–value attention, but measurement-specific spatial attention pooling. We will clarify this terminology and emphasize that its contribution is target-specific spatial aggregation for constructing distinct biometric representations in multi-measurement neonatal biometry.
[R3] Fine/coarse routing threshold. We want to clarify that Sobel complexity map c is normalized per-sample to [0, 1], τ is fixed as the midpoint of [0,1] complexity range. Setting τ = 0.5 intuitively means that locations with normalized edge complexity above 0.5 are preferentially routed toward the fine branch. Routing sharpness is controlled by temperature t, learned end-to-end, allowing the model to adapt how aggressively it differentiates regions. We will clarify the Sobel complexity map normalization, τ justification explicit in Section 2.3.
[R3] Dataset limitation. dHCP is the largest publicly available standardized neonatal MRI cohort for controlled method development. We will state explicitly that multi-site validation remains future work.
We appreciate all reviewers for their efforts.
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.
Given contraversary comments, it would be good to clarify those key concerns from the reviewers.
- 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.
I agree with the majority of the reviewers.
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.
After rebuttal, R1 maintained Accept and R2 remained Weak Accept, while R3 maintained Reject, citing insufficient empirical support for key components such as AAR and limited evidence for the claimed interpretability benefits. While the work addresses an important clinical problem, the current evidence is insufficient to fully support its broader methodological and clinical claims, leading to my recommendation of Reject.
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 authors see merit in the provided rebuttal answers for this paper.
The strengths of a highly relevant topic, well formulated methodology through a two-stage approach, and comprehensive evaluation, are met with remaining concerns of no significant performance gains (with however attention maps provided as a new benefit) for example. All in all, the methodology could result in interesting discussions.