List of Papers Browse by Subject Areas Author List
Abstract
Accurate and comprehensive characterization of brain anatomy across volumetric and surface domains is fundamental to neuroimaging research and clinical practice. Existing methods, however, are often modality-specific and task-restricted, requiring separate models for segmentation, parcellation, or surface reconstruction, thus failing to leverage the intrinsic anatomical complementarity across representations and limiting their utility in heterogeneous clinical settings. In this work, we introduce uBrain, a prompt-driven unified framework that performs simultaneous volumetric segmentation and surface reconstruction within a single model, adaptable to diverse imaging modalities (i.e., T1W, T2W, FLAIR, PD, and CT). Central to our framework is a guided dynamic routing mechanism conditioned on modality- and task-specific prompts, enabling adaptive feature learning and task-oriented decoding across diverse acquisitions. To further enhance generalization under domain shift, we propose a Gaussian Mixture Model-based data synthesis strategy that simulates heterogeneous imaging characteristics. Extensive evaluations on a large-scale cohort of 7,662 subjects demonstrate that uBrain achieves state-of-the-art performance, attaining an average Dice score of 93.21% for volumetric segmentation and an average symmetric surface distance of 0.23 mm for surface reconstruction on unseen external datasets. Moreover, our method maintains robust performance on challenging acquisitions (including thick-slice and ultra-high-field imaging), underscoring its promise for reliable deployment in real-world clinical environments.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1637_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{LiaZif_AVersatile_MICCAI2026,
author = { Lian, Zifeng AND Liu, Jiameng AND Cao, Shui AND Liang, Zhichao AND Shi, Feng AND Shen, Dinggang},
title = { { A Versatile Prompt-Driven Framework for Brain Segmentation, Parcellation, and Surface Reconstruction from Diverse Modalities } },
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 unified model, named as uBrain, to perform multiple brain image analysis tasks, including tissue segmentation, parcellation and cortical surface reconstruction. The model can be applied to both MR and CT images. To develop such as unified model, modality-adaptive encoder and task-driven decoder are introduced, which are integrated within a mixture-of-expert framework.
- 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.It is a neat idea to unify multiple tasks including segmentation, parcellation and cortical surface reconstruction within a single model.
2.As Table 3 shows, the proposed model shows good performance for multiple tasks on both internal and external datasets.
- 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 paper proposes a Mixture of Modality Experts (MoME) model. However, it seems missing closely related works in the literature review, which are also called MoME with similar although not the same idea, including: [1] Xinru Zhang et al. A foundation model for brain lesion segmentation with mixture of modality experts. MICCAI, 2024. [2] Conghao Xiong et al. MoME: Mixture of multimodal experts for cancer survival prediction. MICCAI, 2024.
2.Sec 2.1: “a pre-constructed Look-Up Table derived from real multimodal data in the training set”. How is the look-up table constructed?
3.Sec 2.1: The section is titled “Physics-Informed Data Augmentation”. However, it is not very clear how physics is involved.
4.Sec 2.2: “Guided by these embeddings, the routing layers compute activation probabilities to dynamically select the optimal sub-pathway for feature processing. “ It may be worth elaborating how the activation probabilities are computed. Also, Fig. 4(c) shows that the routing is hard, instead of using soft probabilities.
5.Sec 3.1: “Ground-truth volumetric segmentation labels were generated using [23] and manually refined by experienced radiologists. . . The ground-truth WM surfaces were derived following [25] using these same labels. “ It may be worth clarifying whether the baseline methods such as SynthSeg and Recon-all-clinical use the same ground truth for training/finetuning, or just being used off-the-shelf. If they are used off-the-shelf i. e. being trained using different kind of ground truth from this work, there may be a systematic bias in the resulting segmentation and cortical surfaces from these methods. This may be worth discussing.
- 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?
There is merit in the method design. But the method description and the ground truth of baseline methods may require further clarification.
- Reviewer confidence
Very confident (4)
- [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 rebuttal has addressed my previous concerns.
It is a nice paper with a novel unified pipeline to perform brain image analysis and surface reconstruction tasks, and validated with solid and intensive experiments.
Review #2
- Please describe the contribution of the paper
The paper proposes uBrain, a unified, prompt-driven framework that simultaneously performs volumetric brain segmentation (tissue, hemisphere, ROI parcellation) and cortical surface reconstruction across five imaging modalities (T1W, T2W, FLAIR, PD, CT) within a single mode.
- 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.
(a) Evaluation scale. The evaluation cohort of 7,662 subjects across 10 datasets (7 public + 3 private), spanning five modalities, five age ranges, and both internal and external splits. (b) Modality and Task coverage. The model ables broad range of modalities and tasks. CT brain image segmentation and surface reconstruction is general novel and impactful. (c) Well-designed Evaluations. The paper evaluations shows that proposed model addresses the paper’s goal. The result analysis and visualizations also strengthen the performance claims. (d) Generalizations to UHF data. The model able to result DICE over 90% on the UHF high slice thickness data, while they are explicitly out-of-distribution is novel for clinical development.
- 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.
(a) GMM augmentation concerns. The GMM augmentation is the base idea of SynthSeg (and it’s following works). while the idea of GMM augmentation helps the domain generalization, the model will still struggle if the GMM generalization does not cover the data that we want to inference / test. The paper seems have not clarify if the GMM tissue statics have also been derived from the validation & test subjects in their experiments. Data leakage is a solid concern. (b) Baseline concerns. There have been great efforts of developing unified / foundation model for the brain image analysis, this paper also seems to contributing to this direction of research. Then without these comparisons, the performance may partly reflect GMM data augmentation advantages rather than architectural innovations. (c) Text prompt design. The method is prompt driven, but the text designs are very generic and simple. Further, no ablation exists to inform the reader whether text embedding outperforms simple one-hot encoding or integer category embeddings, which is a natural baseline for the prompt-conditioning claim. t-SNE shows cluster separations, but is not sufficient since high-dimension space separation is easy to achieve. (d) Evaluation metrics. The paper only shows ASSD for surface, average DICE for segmentation. To show the model is ready for clinical utility, more metrics and regional-level numbers are needed.
- 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.
(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 proposed uBrain addresses the fragmentation of brain analysis pipelines across modalities and tasks. The evaluation scale are the strongest aspects of the paper. The ablation study is good and the routing visualizations are interpretable.
However, the paper is weakened by the absence related unified SOTA frameworks, the trivial text prompt design, insufficient metrics. The combination of these weaknesses makes it difficult to attribute the reported gains confidently to the proposed architectural innovations versus training data volume advantages.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
(a) GMM augmentation data leakage. Rebuttal point 2 states the LUT is computed only from the training set, this is largely addressed.
(b) Baseline concerns. The available results still does not show whether without GMM, proposed model is better than the Synthseg & Recon-all-clinical baselines (Table 2 & Table 3), since Table 3 only shows the comparison between full model and ablation w/o GMM. Table 2 and Table 3 numbers are not directly comparable (table 2 are individual modality full dice & full ASSD results, table 3 are modality average results of Tissue, ROIs and different surfaces). And more related model comparison should be added to the research to improve the broader claim. The concern of too few baseline could not be addressed based on the policy, which results this as a legit limitation.
(c) Text prompt design. “better than traditional hard-coded categorical labels” remains unsupported and should be softened or the ablation deferred to a limitations statement.
(d) Evaluation metrics. Author agree to this limitation, and under the no new results rule, based only on the rebuttal text, this is unaddressed. This downgrades the paper contribution credibility. For example, in terms of Cortical Surface Reconstruction task, the self-intersection rate is critical and distill to if the surface reconstructed is useable in downstream analysis. This intersection rate could be high even when ASSD metric being low. As such, the absence of detailed metrics and results is a limitation.
The current paper form is not ready to cover authors’ claims in my opinion.
Review #3
- Please describe the contribution of the paper
This paper presents BRAIN, a prompt-driven framework that performs volumetric segmentation and surface reconstruction within a single model across MR and CT imaging modalities. The general premise is that volumetric and surface segmentation have traditionally been treated as separate problems.
- 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 architecture is clearly presented.
- 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.
This reviewer finds some difficulty with this premise. From early work in cortical processing pipelines, these tasks have been tightly coupled, and more recent joint learning approaches have emerged, including Vox2Cortex and SegRecon. While these methods are largely MR-centric or modality-specific, the broader claim that this has not been considered as a dual problem is somewhat problematic. A pre-trained text encoder is used for prompt-based inputs; however, it is unclear how these prompts function in practice. Specifically, it is not evident whether they are truly embedded as free-form text or instead operate more like a constrained set of tokens, as the range of possible inputs appears limited. The provenance of the ground truth is also unclear. The annotations appear to be algorithmically generated, but it is not specified how much manual radiological editing or validation was performed. This raises concerns regarding comparisons with state-of-the-art methods, which may have been trained using different sources of ground truth or alternative annotation procedures. As a result, direct performance comparisons are difficult to interpret.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The visualizations are clear and compelling, but the generalizability of the approach remains uncertain. In addition, the level of implementation detail is insufficient to support reproducibility, and no code or external resources are provided for further evaluation.
- 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?
Overall, this is an interesting pipeline that demonstrates the ability to learn a specific representation effectively. However, there are challenges in understanding how the ground truth is defined, how well the method generalizes, and what constitutes an appropriate baseline. This is particularly important in the cross-modality setting. Without a detailed ablation study, it is difficult to determine which components of the model are most critical to performance. A key concern is that the baseline methods may have been trained toward different definitions of truth, making comparisons less meaningful.
- Reviewer confidence
Very confident (4)
- [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.
My enthusiasm is increased with the rebuttal. I see that the reproducibility concerns have been mitigated someone. It still remains a weak-accept, but given two choices, I’ll lean towards accept.
Author Feedback
We thank the AC and all reviewers for their constructive feedback.
1.Ground-Truth Curation & Baseline Fairness (R1, R3) We clarify that our annotations are not raw algorithmic outputs. Initial volumetric labels were generated using Ref. 23, followed by rigorous visual QC and correction by 20 trained raters using ITK-SNAP. Cases with extensive segmentation errors, abnormal intensity distributions, or severe artifacts were excluded. Corrected cases took ~20 minutes each, with ~1–3% of voxels modified. Final labels were approved by 8 experienced radiologists. Overall, ~3–5% of cases were excluded, leaving 7,662 scans. Surface targets were derived from these labels following Ref. 25.While off-the-shelf baselines (SynthSeg and Recon-all-clinical) were not retrained, the comparison remains fair: both our annotations and the baselines are built on FreeSurfer anatomical conventions. For the adult data used in our study, Ref. 23 uses FreeSurfer-derived QC labels, while the baselines are trained on FreeSurfer-related labels. Thus, all methods are evaluated under a common anatomical convention and a single reference standard.
2.LUT Construction & Leakage Control & “Physics-Informed” Augmentation (R1, R2) For each training scan, we compute class-wise intensity statistics, i.e., mean mu_k and standard deviation sigma_k, from voxels of tissue class k and store them in the LUT. During augmentation, for a source image I_src, we compute source statistics, sample target statistics of the same class from the LUT, and remap voxels with L(v)=k as described in Sec. 2.1.To prevent data leakage, the LUT is computed only from the training set. In our work, “Physics-Informed” means simulating structured acquisition shifts (LUT-based contrast remapping, random bias fields, Gamma transforms, and partial-volume effects) rather than arbitrary perturbations.
3.Prompt Conditioning & Hard Routing (R1, R2, R3) uBrain uses controlled prompts for predefined medical concepts, rather than free-form text, to provide stable anchors for modality-adaptive encoding (MAE) and task-driven decoding (TDD). Unlike orthogonal one-hot vectors, text embeddings capture semantic correlations across modalities and tasks (Fig. 4). Modality embeddings guide soft MoME expert aggregation through cross-attention, whereas task embeddings drive hard Automatic Pathway (AP) routing. For each AP module l, the router produces logits z_l = W_l p_task + b_l over M=5 paths, and softmax(z_l) gives intermediate activation probabilities. During training, a hard Gumbel-Softmax with temperature tau=1.0 produces one-hot path selection while preserving differentiability through the straight-through estimator. During inference, we select the maximum-probability path by argmax. Thus, Fig. 4(c) correctly shows hard path selection, not soft mixing.
4.Novelty & Baselines (R1, R2, R3) We will cite and discuss Vox2Cortex, SegRecon, and prior MoME methods. Unlike these works, which primarily target MRI surface reconstruction, with an auxiliary volumetric component often limited to simple filled-WM segmentation, or focus on lesion segmentation, uBrain uses MAE and TDD in one prompt-driven model to unify multimodal inputs and multiple co-equal anatomical outputs, rather than merely coupling volume and surface. Table 3 shows that MAE+TDD alone improves over SwinUNETR baseline before GMM augmentation, supporting the architectural efficacy of MAE and TDD. Since no single foundation model matches our broad input-output scope, we compare against strong task-specific SOTA baselines (SynthSeg and Recon-all-clinical).
5.Evaluation Metrics (R2) We agree regional metrics are vital. Due to space limits, we reported standard average metrics (Dice, ASSD). The revised Supplementary Material will include detailed ROI-level metrics (e.g., hippocampus, ventricles) and HD95.Our current results demonstrate broad multimodal potential as a unified framework, with rigorous clinical deployment validation as future work.
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 reviewers generally acknowledge the merit in proposing a unified, prompt-driven framework for brain image analysis across multiple modalities and tasks. However, significant concerns were raised across all reviews, focusing on the novelty of the unified approach in light of prior work, the clarity and justification of certain methodological components, and the issue of baseline comparison given the unique ground truth generation. Specifically, the lack of comparison against other recent unified/foundation models and the insufficient detail for reproducibility weaken the paper’s claims.
- 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 rebuttal successfully resolves the concerns regarding data leakage and ground-truth definitions. I recommend accepting this paper.
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.
I want to start acknowledging that I agree with the reviewers in highlighting that merging tasks is a worthwhile endeavour and a great strength of the paper. Similarly, I can understand why two of the reviewers decided to accept after rebuttal as some concerns have been successfully addressed and I also agree that a large-scale evaluation is a great strength in itself.
Having said that, I cannot help but side with reviewer #2 in terms of major concerns not addressed. My decision is made on the basis of considering this submission a methodological paper and expecting a few different things from it:
- Methodological rigour. Having read the paper and rebuttal, I am unconvinced of the following claim: “Unlike orthogonal one-hot vectors, text embeddings capture semantic correlations across modalities and tasks”. Figure 4 does show that the embeddings are clustered but it does not provide evidence on whether single categorical variables, passed through an encoder would be equivalent. Similarly, “‘physics-Informed’ means simulating structured acquisition shifts” is a fairly loose interpretation of what physics-informed means and misleading. If data augmentation was truly “physics-informed”, I would have expected that the authors would exploit properties from magnetic resonance and x-ray tomography. Furthermore, whether linear scaling based on Gaussian distributions is novel is arguable, too. The same could be said for most of the methodological contributions beyond unifying segmentation and cortical reconstruction.
- A comprehensive and rigorous evaluation. I do agree with the authors that a rebuttal is not the place to address new experiments, but the lack of proper comparisons in the original manuscript is a major concern, nonetheless. With this I am echoing reviewer #2’s complaints about baseline concerns. Moreover, I also agree that not addressing crucial evaluation metrics for downstream tasks (even if it could be due to rebuttal policy) downgrades the author’s contributions. Evaluation, in general, should cover relevant metrics to understand the applicability of a new method. Limiting the evaluation to metrics that do not provide the whole picture is still a major concern, whether a rebuttal allows for it or not.
Meta-review #3
- 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.
Given mixed opinions, I carefully reviewed the paper and rebuttal. While novelty concerns are mostly addressed, I still oppose acceptance for the following reasons:
1.Missing baselines: I highly disagree with the author’s response that only one baseline is relevant for comparison. Comparisons to the following methods would have been highly relevant and are currently missing to judge the presented work against the state-of-the-art: SuperSynth (contrast-agnostic multi-task model), PhySeg [a] and GOUHFI [b] (contrast-agnostic segmentation), Anatomix [c] and Brain-ID [d] (contrast-agnostic brain MRI foundation models).
2.Potentially biased results: the only baseline was not even retrained on the authors’ data. Using a common list of labels is not a safeguard against annotator bias (human or automated), which often have different ways of segmenting the same regions. Right now, it is unclear whether the reported gains come from methodological improvements or labelling differences.
3.Unclear methods/limited reproducibility: as noted by reviewers, the method description remains largely superficial and omits key information about the main contributions (modality MoME, dynamic pathway rerouting). Combined with the lack of code release, this makes this work not reproducible.
4.Insufficient literature review: for MICCAI, even a concise literature review should be conducted. Here, decades of previous works are reduced to two lines citing all learning-based segmentation works together [9–16].
5.The augmentation pipeline after GMM sampling (partial-volume, bias field, gamma transform) is directly taken from SynthSeg and should be explicitly acknowledged. Overall, the work is promising, but these major issues make it unsuitable for acceptance at MICCAI in its current form.
[a] Borges et al., MedIA, 2024 [b] Fortin et al., Imaging Neuroscience, 2025 [c] Dey et al., ICLR, 2025 [d] Liu et al., ECCV, 2024