List of Papers Browse by Subject Areas Author List
Abstract
Accurately characterizing brain morphological changes throughout human lifespan is crucial for understanding brain development, aging, and disorders. At the core of this endeavor lies cortical surface reconstruction (CSR), which underpins the computation of essential brain morphological features. How-ever, existing CSR methods face two major limitations. First, cortical surfac-es are typically reconstructed from 3D MRI data with high isotropic resolu-tion, which is confined to research settings. In contrast, clinical MRI scans are collected with high in-plane but low through-plane resolution. Second, most CSR pipelines are designed either for adult or pediatric populations, re-stricting their applicability across the lifespan. To this end, we develop a deep learning framework that harnesses MRI super-resolution (SR) as a bridging mechanism, leveraging the complementary information SR provides to jointly perform SR and CSR with a coarse-to-fine strategy. Specifically, we introduce a dual-decoder age-conditioned temporal attention network (DATAN) with a shared encoder, which simultaneously performs CSR and SR from thick-slice clinical MRI. By jointly training on the SR task, the shared encoder captures richer cortical features, thereby enhancing CSR per-formance. Through a two-stage coarse-to-fine approach, incremental refine-ments in the SR output progressively restore fine-scale details otherwise lost in low-resolution scans, ultimately improving CSR fidelity. Furthermore, to facilitate accurate CSR across the lifespan, we exploit the age-conditioning module of our framework and train our model on a large, diverse MRI da-taset spanning ages from 1 to 100 years. Experimental results demonstrate that our method, despite requiring only thick-slice clinical MRI scans, achieves consistently improved CSR performance across the entire human lifespan.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1500_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{DonXiu_Lifespan_MICCAI2025,
author = { Dong, Xiuyu and Tang, Kaibo and Hu, Dan and Wu, Zhengwang and Wang, Li and Lin, Weili and Li, Gang},
title = { { Lifespan Cortical Surface Reconstruction from Thick-Slice Clinical MRI } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15961},
month = {September},
page = {302 -- 312}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a coarse-to-fine deep learning framework for cortical surface reconstruction, incorporating age as an additional input to enable lifespan modeling. The approach jointly performs super-resolution of low-resolution MRI data and predicts the corresponding cortical surface.
- 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 jointly performs super-resolution of low-resolution clinical MRI data and reconstructs the cortical surface, while incorporating the subject’s age as an additional factor in the modeling process.
- 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 methodological novelty is somewhat limited, as the coarse-to-fine reconstruction strategy has already been introduced in previous works such as CorticalFlow++. Moreover, since the paper highlights its contribution to lifespan modeling, the experiments should specifically present results demonstrating improvements in both low-age and high-age subjects to better support this claim.
- 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.
- 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/2025/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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposes an age-aware deep learning framework for cortical surface reconstruction. However, the methodological novelty is somewhat limited. Additionally, the experimental design is not sufficiently convincing to demonstrate the practical value of the proposed method. In particular, there are no strong results supporting improvements on subjects at the extremes of the age spectrum. Furthermore, in practice, it is common to first upsample low-resolution data to isotropic super-resolution using existing methods, and then perform cortical surface reconstruction using state-of-the-art techniques, which is not adequately compared against in this work.
- 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 authors have clarified the novelty of their approach and its key differences from state-of-the-art methods. I believe this work has potential applications in clinical settings, particularly in simplifying processing pipelines, as it can handle anisotropic and low-resolution data. Despite its limitations, I find the overall contribution valuable and recommend accepting the paper.
Review #2
- Please describe the contribution of the paper
This work introduces a two-stage deep learning framework for cortical surface reconstruction (CSR) from simulated thick-slice MRIs. It leverages a dual-decoder age-conditioned temporal attention network (DATAN) that jointly performs super-resolution and CSR. Large-scale testing is performed, and results compared to benchmark methods look compelling.
- 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 is well explained and the rationale behind each module is properly given. Experiments, including ablation studies, are well conducted for the most part and spanned over several 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.
Despite its potential, the study falls short of addressing real clinical variability. All experiments are based on isotropic datasets with simulated thick-slice images achieved by axial downsampling by a factor of 4, which may not capture the full spectrum of clinical acquisition settings and artifacts.
The aggregated performance metrics in Table 2, though impressive, obscure important subtleties — for instance, how model performance varies across different age ranges and datasets. Lastly, while external validation on ADNI is provided, the rationale for not testing baseline methods on the same dataset is unclear.
Minor comments:
-
HCP matrix size 260 x 311 x 26 looks incorrect. Are you sure it’s not 260 for the last dimension?
-
There are a lot of (linear) interpolation steps in the preprocessing and in the proposed pipeline. Do the authors have an understanding of how much this is influencing the final results, considering that cortical measurements need to be sub-millimeter accurate?
-
It would be nice to report the performance of the proposed and benchmark methods vs. iBEAT and FreeSurfer separately.
-
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is well-written, the method is novel, and the experiments encompass several datasets with compelling results. However, the lack of real clinical data and the simplistic thick-slice simulation limit the scope of the paper to non-clinical scenarios.
- 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.
I appreciate the author’s response, particularly regarding the method’s performance in age-specific experiments. While I still believe the clinical experiments are somewhat simplistic and capture only a limited portion of the diversity present in clinical data, the paper represents a valuable contribution to the field.
Review #3
- Please describe the contribution of the paper
This paper presents a novel neural network architecture, DATAN, which addresses cortical surface reconstruction (CSR) and MRI super-resolution (SR) from low-resolution clinical MRI data. The proposed two-stage framework is applied to cortical surface reconstruction for MRI data across the human lifespan, demonstrating method’s generalizability across a wide range of age groups.
- 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 proposes a multi-stage strategy for surface reconstruction from thick-slice, low-resolution clinical MRI data, which has potential applicability in real-world clinical settings.
2 . The method is trained and tested on a large, diverse multi-site dataset spanning a wide age range.
- 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.
- Table 3 shows high variance in SIF, particularly for white matter surfaces, raising concerns about the reliability of the reported improvements. Additionally, pial surface results between Stage 1 and Stage 2 are marginal.
- The evaluation setup excludes common clinical practices like using external SR models before CSR, making the comparison with LR-only baselines potentially unfair.
- Despite training on a wide age range, no age-specific results are reported. The paper lacks analysis on whether age-related cortical features are preserved, weakening claims about lifespan generalization.
- The inclusion of Laplacian and normal consistency losses lacks justification or ablation.
- Addition to geometric accuracy, clinical metrics like cortical thickness should be evaluated to assess downstream utility.
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The evaluation is broad in scope and demonstrates the method’s potential across diverse datasets. However, the ablation study results are not strong enough to fully validate the two-stage design.
- 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
Author Feedback
We thank all reviewers for their constructive comments and recognition of our contributions: “jointly performs cortical surface reconstruction (CSR) and super-resolution (SR) from low-resolution (LR) clinical MRI data” (R1, R3, R4). Evaluation (R1, R3, R4): Regarding age-specific results, we have reported in Figs. 2 and 3. We extensively evaluate our method on a large-scale T1w MRI dataset composed of 6 public and 1 in-house datasets, each corresponding to a distinct age range spanning infancy to late adulthood (Table 1). To showcase our method’s generalizability across age groups, we visualize reconstructed cortical surfaces for representative subjects from each age range (dataset) and their corresponding error distance maps (Figs. 2 and 3) for clearer spatial interpretation, instead of reporting summarized age-specific quantitative metrics. Our method performs particularly well in younger age groups, whereas baseline methods struggle to capture fine cortical structures due to large developmental variability. Regarding comparison with iBEAT and FreeSurfer, we primarily focus on learning–based approaches, as our framework is also learning-driven. In addition, despite being widely used, iBEAT and FreeSurfer are computationally intensive, especially for large-scale processing. Nevertheless, we tested several representative subjects with both tools and observed notably inferior performance, especially in young ages. Regarding comparison with LR input, our model is specifically designed to reconstruct cortical surfaces directly from LR MRI. The purpose of the SR task is to encourage learning of richer feature representations in cortical regions. In this setup, SR is not an explicit pre-processing step but rather an implicit byproduct of the joint learning process. Therefore, we use only LR inputs for fair comparison. Clarification (R3, R4): Our model has fundamental differences from CorticalFlow++ (CF++). First, in CF++, “coarse-to-fine” refers to the use of a series of deformation blocks to successively refine reconstructed surfaces, whereas in our work, it refers to the two-stage refinement strategy which improves CSR performance on LR lifespan MRIs (Table 3). Second, CF++ is designed for isotropic high-resolution MRI, whereas our method is specifically tailored to LR, thick-slice clinical MRI input. Most importantly, the innovation of our work lies in the shared robust feature representation learned through training jointly on SR and CSR. Specifically, our SR task significantly enhances CSR performance on LR lifespan MRIs (Table 3), as the SR decoder forces the shared encoder to learn richer and more robust feature representations. This enables us to reconstruct cortical surfaces from LR MRI, addressing a major gap in prior research. Furthermore, our method is generalizable across ages, as opposed to previous methods, which are typically restricted to certain age groups. We will clarify these in the revision. We apologize for not clearly explaining the rationale behind Laplacian and normal consistency losses. These losses are well-established in prior works [14,19], which show their effectiveness as regularization terms for CSR. Hence, we adopt them as default in our loss function. We will clarify it in the revision. Clinical Variability (R1): In our current study, we follow prior SR works (e.g., Zhao et al., TMI 2020; Wu et al., MICCAI 2024) by simulating thick-slice clinical MRI through axial downsampling, motivated by the prevalence of anisotropic spacing in clinical scans. With our current straightforward simulation, we can achieve substantial improvement over baseline, demonstrating the superior effectiveness of our framework. We acknowledge that clinical MRI variability can extend beyond slice thickness. We thank the reviewer for raising this point and have already begun addressing it by incorporating more diverse data augmentation, e.g., variable downsampling in random axes, noise, motion artifacts, etc.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
All reviewers find merit to this paper and recommend acceptance. That said, the way low-resolution images are generated is extremely restricted and simplistic, and I doubt very much that the proposed method works on actual clinical scans. The absence of an evaluation (even qualitatively) on real clinical scans makes it impossible for the authors to convince me otherwise. All test images are also “fake LR”, whose contrast and slice profile exactly match those of the (also fake) training set. The comparison with baselines that assume ~1mm input resolution is therefore extremely biased. I do not find that the authors rebuttal addresses these major issues. Had I been a reviewer, I would have (strongly) recommended rejection.