List of Papers Browse by Subject Areas Author List
Abstract
Early diagnosis and quantitative assessment of pediatric hydrocephalus are critical, yet automated segmentation remains challenging due to severe anatomical deformation and substantial appearance variation across early neurodevelopment, imaging modalities, and resolutions. Current deep learning approaches are often limited to ventricle-only segmentation and tailored to specific acquisition protocols, hindering their clinical utility for comprehensive joint tissue analysis and longitudinal evaluation. We propose SynHydro, a novel biomechanics-driven simulation and domain-randomized augmentation framework. SynHydro synthesizes diverse hydrocephalus anatomies from readily available healthy tissue labels and generates corresponding MRIs with randomized image characteristics. This creates a large-scale, synthetic training dataset encompassing a wide spectrum of disease severity, MRI contrast, noise, resolution, and artifacts for joint tissue-ventricle segmentation. Extensive experiments on in-house and public datasets demonstrate that a segmentation model trained exclusively on this pure synthetic data achieves performance comparable to models trained with real hydrocephalus annotations. More importantly, SynHydro-trained models exhibit superior cross-modality and cross-resolution robustness, and reliably preserve pre-/post-operative volumetric trends. By eliminating the dependency on labor-intensive pathological annotations, SynHydro provides a powerful and practical solution for building robust, generalizable segmentation tools for pediatric hydrocephalus monitoring across the clinical imaging landscape.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1217_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{RenZeh_SynHydro_MICCAI2026,
author = { Ren, Zehua AND Zhai, Jingyan AND Wang, Haifeng AND Bo, Yixuan AND Wang, Fan AND Zhao, Cailei AND Lian, Chunfeng},
title = { { SynHydro: Biomechanics-Driven Domain Randomization for Hydrocephalus-Agnostic, Generalizable Tissue-Ventricle Segmentation Across Age, Modality, and Resolution } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
this work proposed a model to generate simulated hydrocephalus cases, and the performed the following tasks on it.
- 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 most interesting idea is to proposed a mechanical driven deformation model to simulate the hydrocephalus. The similar idea has been used in other brain development research but to generate a whole brain is still interesting
- 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 primary concern is how to evaluate the accuracy of the proposed synthetic model. The author didn’t provide any data or experiments to evaluate this part. It is high required and a critical part of this work. 2.the following task is not very well designed, for example only using the synthetic model to train the NN-UNET.
- 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 mechanical model of the brain deformation
- 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
Review #2
- Please describe the contribution of the paper
This paper proposes a three stage method for robust and accurate pediatric image segmentation with hydrocephalus, using only healthy images. The method comprises of three stages: (1) creation of realistic hydrocephalus maps with bio mechanical deformations from healthy label maps, (2) creation of images with domain randomization, and (3) training a segmentation model using both normal and the created synthetic dataset. The paper claims to segment hydrocephalus images robustly among different modalities, ages and resolutions.
- 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 presents a simple method to utilize healthy images to create a model that is robust to large anomalies in the brain. -The writing is clear and understandable. -The idea of synthesizing images with hydrocephalus anatomy with the described bio mechanical driven method is novel, and performing well (Dice score of ~0.85) to segment white matter, gray matter and ventricles in pediatric T1w MRI. -
- 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.
- “Real-Augmented: Trained on 1400 augmented data synthesized from all available training labels (both longitudinal and control), representing a strong data-augmentation baseline with real pathological labels”, “Real-Augmented achieves the best performance on T1w, serving as an approximate upper bound” The usage of baseline and upper bound interchangeably is confusing. Could they consider changing this description?
- What is the reason for selecting only T1w sequences for testing quantitatively? The results lack the cross modality segmentation robustness.
- In datasets, the authors mention they have 132 longitudinal sessions and 131 control scans. Then in the implementation section, they mention having 125 longitudinal and 47 control 3D images. How did they select these subjects for annotating, is it random?
- HyKid dataset has ground truth segmentations available for gray matter, white matter, cerebrospinal fluid, the choroid plexus, and the lateral ventricles. What do the authors imply when they say “ground-truth labels consistent with the purpose of this study is not available” ?
- Both the qualitative and quantitative results show similar performance to the “Real augmented” case. The augmentations seem to capture modality differences & the resolution difference and seem to be useful looking at the qualitative improvement from real paired to real augmented. This generalization raises the question of applying augmentations to the baseline (control only) model. I believe the setting with strong spatial and intensity augmentation to control set, without explicitly creating the hydrocephalus images should be included in the results as well for a fair comparison.
- The results lack comparisons with segmentation methods similarly trained with domain randomization and have shown good performance over various anatomical changes [1,2]. The authors have cited the first paper yet have not included any results. I believe it would be useful to compare with the existing methods, as they are known to not fail in cases of such large enlargement of ventricles (see the results in a similar work dealing with abnormalities [3]).
- The title and the motivation suggests that the method is age, modality and resolution. I do not think the results are supportive of this claim. It would be useful to show results on different modalities, ages and resolutions separately to be able to claim this. Figure 3 if helpful for understanding the consistency over modalities yet it just shows “a” representative subject, which is not convincing. Table 2 shows correlation among different modalities, but without GT labels on different modalities it’s hard to judge. I could suggest using a tool like SynthSR [4] to be able to map all sequences and resolutions into the same space for evaluation purposes, and then you can compare your method with those obtained labels for example.
- The proposed model’s advantage is not using the labeled images with pathologies. To highlight the need for the method, it could be useful to do the ablation study at different number of labeled subjects with the pathology (i.e. does 50% control, 50% pathological subjects in the real augmented setting perform considerably worse than the real augmented upper bound presented in results?).
[1] B Billot, DN Greve, O Puonti, A Thielscher, K Van Leemput, B Fischl, AV Dalca, JE Iglesias. SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis, 83, 102789 (2023). [2] B. Billot, C. Magdamo, Y. Cheng, S.E. Arnold, S. Das, & J.E. Iglesias, Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets, Proc. Natl. Acad. Sci. U.S.A. 120 (9) e2216399120, https://doi.org/10.1073/pnas.2216399120 (2023). [3] Z. Shang et al. Towards contrast- and pathology-agnostic clinical fetal brain MRI segmentation using SynthSeg, NeuroImage, Volume 327, 2026, 121729, ISSN 1053-8119, https://doi.org/10.1016/j.neuroimage.2026.121729.[4] JE Iglesias, B Billot, Y Balbastre, C Magdamo, S Arnold, S Das, B Edlow, D Alexander, P Golland, B Fischl. SynthSR: a public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science Advances, 9(5), eadd3607 (2023).
- 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 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.
(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?
The idea of synthesizing images for specific structural changes in brain is a useful idea and the authors show a method that shows good performance for segmenting ventricle, gray matter and white matter structures under modality and resolution changes. However, the results lack comparison with very similar methods, a trivial baseline of control images with train-time augmentation, the appropriate semi supervised ablation and the evaluation on different sequences, ages and resolutions (unlike what’s claimed). Quantitative evaluation of one of two datasets is missing.
- 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.
I would like to thank the authors for responding to my comments and taking them into consideration for the next iteration of their paper. Thank you for your clarifications about the dataset, and the exact training settings. From my understanding, the authors use augmentation and domain randomization interchangably, which was unclear for me before. Thanks for clarifying the “real augmented” as a comparable method and not an upper bound, this clarification renders the paragraph on page 7 outdated, please consider changing this. The following concerns of mine, however; were not resolved:
- Authors are claiming their label protocol is specific in three of the comments, but from the description and the results table, both the original HyKid labels and the results have ventricle, gray matter and white matter classes that are comparable without bilaterality.
- Correlation results on consistency seems to be aggregated over multiple resolutions and sequences, without explicit explanations on how. The numbers reported on Table 2, hence, are hard to understand in terms of performance.
- SynthSeg like methods are relevant, and they are not corresponding to authors’ normal only setting. They use labels other than control scans for domain randomization, and can be used to obtain segmentations for multiple resolutions and sequences. I don’t think that they could be excluded from (at least) qualitative results or thought outside the scope of this work. -The main contribution is the ability of generating real-like labels when the pathological labels are scarce, but there are no comments or results presented about this tradeoff (ablating ratio of control vs pathological in the real augmented setting).
I think the contribution of creating a diverse and controllable hydrocephalus deformation is valuable and the results show that the new generated labels are almost as useful as using the real pathological labels. As the authors also commented on the feedback, complete evaluation is not yet there. In my opinion, eliminating the label protocol mismatch by merging/reassigning labels and showing that “generation of the anatomically plausible pathology is useful and necessary compared to highly diverse synthetic data generation from various source labels (like SynthSeg)” would make the results much sharper and convincing.
Review #3
- Please describe the contribution of the paper
This paper introduced SynHydro, an end-to-end biomechanicsdriven simulation and domain randomization framework for robust, hydrocephalusagnostic segmentation.
- 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.
This paper introduced an end-to-end biomechanicsdriven simulation and domain randomization framework and creates a large-scale, synthetic training dataset. Extensive experiments on demonstrate that a segmentation model trained on this synthetic data achieves performance comparable to models trained with real hydrocephalus annotations.
- 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.
There are several points in this paper need clarification: 1: The authors propose a biomechanical-driven hydrocephalus anatomy synthesis process. Is there any relevant theoretical support for this process, and can previous studies confirm that such modeling approach accurately reflects the actual progression of hydrocephalus? 2: The authors state: “Given a label map (original healthy or simulated hydrocephalus), we synthesize paired data with a domain-randomization strategy, following label-to-image paradigms [1,5,6,4].” It is recommended that the authors present examples of the synthesized label maps and the corresponding multi-contrast MRI images generated from these maps, along with a comparison to real MRI images from hydrocephalus patients. 3: Please indicate the specific modality (e.g., T1, T2) for each column in Figure 2.Additionally, clarify whether the “input” row represents images from the authors’ constructed dataset or the original images. 4: The authors mention that “the image is downsampled by a factor of 3-7 to create an anisotropic low-resolution (LR) volume.” Is this factor based on empirical experience or derived from literature? Furthermore, are the generated low-resolution data used in constructing the SynHydro dataset? 5: The training parameters of the model should be described in detail, including batch size, learning rate, and the input 3D image patch size. 6: Does the “Real-paired” training data consist of the 125 (longitudinal) and 47 (control) 3D T1w image-label pairs? 7: It is advisable to elaborate on the differences between SynHydro and the other three baselines. For instance, regarding the Baseline (Normal-Only) method, is the 1400-subject dataset obtained in a manner consistent with SynHydro but sampled from 35 healthy controls, or is it generated using conventional data augmentation techniques such as flipping and rotation? 8: If the datasets for other baselines are generated using conventional data augmentation methods and have the same dataset size, mixing a certain proportion of real healthy control scans and hydrocephalus patient scans with data augmentation might achieve better model performance than the proposed SynHydro method (as shown in Table 1). If so, what is the clinical significance of the proposed framework? 9: The authors select a 3D nnU-Net model as the segmentation model. Does the model take the patient’s full head scan as input? Additionally, is the proposed SynHydro framework also implemented in 3D?
- 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 authors propose a biomechanics-driven simulation and domain-randomized augmentation framework and create a large-scale synthetic training dataset. However, the clinical authenticity of the synthetic data and the theoretical support for the framework have not been sufficiently validated.
- 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’ original work is innovative, and the rebuttal has addressed my relevant concerns.
Author Feedback
We thank all reviewers for their constructive comments. We will revise the manuscript to clarify the following points.
[R1-Q1, R3-Q1] SynHydro is not a patient-specific biomechanical solver or a model for predicting exact hydrocephalus progression. Each synthetic case is generated from a healthy label map, so no paired real counterpart exists for voxel-wise validation. Our goal is to create diverse, controllable, anatomically plausible hydrocephalus-like deformations. Thus, validation should focus on whether synthetic data provide useful pathological supervision rather than exact mechanical accuracy. This is supported by T1w DSC, qualitative results, and volumetric consistency on unseen real data. We will clarify this and add clearer synthetic examples.
[R1-Q2, R2-Q5, R3-Q7-Q9] The contribution is not simply training nnU-Net with synthetic data. We use 3D nnU-Net as a fixed backbone to compare supervision sources under the same protocol. Baseline (Normal-Only) uses healthy/control labels with the same label-to-image domain randomization, matching R2’s requested strong control-only augmentation baseline. SynHydro adds hydrocephalus-specific deformation to these labels. Real-Augmented uses real annotated healthy and hydrocephalus labels with the same synthesis pipeline. Real-Paired directly uses real 3D T1w pairs. These settings isolate generic domain randomization, explicit pathological simulation, real pathological labels, and direct real-image supervision. We will clarify that all methods operate on 3D volumes.
[R2-Q1] “Real-Augmented serves as an approximate upper bound” was inaccurate. The intended in-domain T1w upper-bound reference is Real-Paired, because it is trained directly on real annotated 3D T1w pairs. Real-Augmented is a strong real-label augmented comparator. We will correct this.
[R2-Q2, R2-Q7] Dice-based evaluation requires complete manual labels. In our dataset, radiologist-refined labels are available only for 3D T1w scans; other modalities and LR scans do not yet have complete labels. Thus, Table 1 reports DSC only on annotated T1w scans. Cross-modality/resolution robustness is assessed through qualitative visualization, volumetric consistency, and longitudinal trends. We agree these are not full voxel-wise multi-modality DSC, and will soften the claim to “promising robustness under heterogeneous age, modality, and resolution conditions.” More complete evaluation is ongoing.
[R2-Q3, R2-Q4, R3-Q6] The 132 longitudinal sessions and 131 control scans refer to the full in-house multi-modal collection, while 125 longitudinal and 47 control cases refer to annotated 3D T1w pairs. The annotated pool was determined by usable 3D T1w acquisition, image quality, successful initialization, and radiologist refinement. A subject-level split avoids leakage, and Real-Paired uses only training cases. For HyKid, labels exist but are not compatible with our 8-class bilateral protocol; direct DSC could be misleading. We are working on label alignment and will provide our best SynHydro predictions.
[R2-Q6, R2-Q8] SynthSeg-like methods are relevant because they also use label-to-image domain randomization. Our Baseline (Normal-Only) is a controlled generic domain-randomization baseline under our label protocol. Its gap from SynHydro shows that intensity/contrast/resolution randomization alone is insufficient for severe hydrocephalus; explicit pathological expansion is important. Direct comparison is complicated by label-protocol mismatch. Preliminary partial-label experiments suggest that limited pathological supervision improves performance and then tends to saturate. We will discuss these points as future/extended work.
[R3-Q2-Q5] We will revise figures to show synthetic labels/MRIs and real hydrocephalus MRIs; indicate modality/resolution in Fig. 2; and clarify that “Input” contains original real scans. The 3–7× downsampling factor was empirically chosen from clinical resolution variation and used in training.
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 reviews are conflicting, it would be good to clarify the 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.
Reject
- Please justify your recommendation.
The reviewers agree that the paper has clinical and techincal merits. However, evaluationsI is not sufficient enough.
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.
While Reviewer 2 maintained a reject score due to evaluation scope, the authors’ rebuttal successfully clarified the dataset details and the specific role of the synthetic simulation framework, leading Reviewer 3 to upgrade their rating to an Accept. Given that the framework provides valuable anatomical supervision where real pathological data is scarce, the paper is recommended for acceptance.
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.
First of all, I discarded R1’s and R3’s main reasons to reject the paper, since they failed to understand that the value of this work is not in the “realism” of the generated images, but in their capacity to enhance the generalisation of downstream models (see SynthSeg and domain randomisation literature). Thus, evaluating the quality/realism of the generated images is not relevant.
In general, I find that the rebuttal successfully addressed concerns about the contributions/positioning of this work on biomechanics model + domain randomisation for robust hydrocephalus segmentation. It also provided much needed clarifications about the experimental set-up and the training frameworks of the tested baselines. Moreover, while R2 is right that Normal-Only is not exactly equivalent to SynthSeg (GMM parameters sampled from multi-modal Gaussians instead of a uniform distribution), I believe this difference to be negligible given the large number of modes (ie contrasts) in this distribution. I also agree with the authors that the direct comparison with the original SynthSeg model requested by R2 is not really relevant, as SynthSeg was shown to provide mixed results on younger populations.
However, I do agree with R2 that quantitative evaluations on multi contrast and multi resolution is missing. I also agree with R2 that the authors could have defined some common labelling between their annotations and the HyKid’s one to at least provide quantitative scores. Right now the claims about robustness to contrast and resolution are only weakly supported by visual evidence only. Personally, I also regret the total lack of statistical testing from the experiments.
Overall, this paper brings a novel and interesting contribution to pathology simulation in brain MRI (the biomechanics deformation model), but is hindered by experiments that do not fully support all the claims. Thus, this paper is borderline accept/reject. Here, I decide to recommend acceptance given the recent surge in papers about domain randomisation for segmentation of brain MRI in younger populations. The main contribution on the deformation model will also foster discussions in the broader geometrical learning community.