Abstract

Cerebrovascular modeling remains a foundational challenge in medical image analysis due to the intricate, multi-scale topology and extreme anatomical variability of the brain’s vascular tree. Traditional discrete representations, such as voxel grids, face challenges in capturing the continuous nature of these highly variable structures, often leading to spatial wash-out during population averaging. In this study, we introduce VITAL, an implicit neural atlas framework for brain vessel modeling. To address the multi-scale nature of the cerebrovasculature, spanning from major arteries to distal branches, VITAL utilizes a dynamically routed Mixture-of-Experts architecture that decouples multi-scale features, enabling effective and continuous vascular reconstruction. To our knowledge, we present the first continuous conditioned probabilistic neural atlas of the human cerebrovasculature, enabling cross-sectional modeling of age-related morphological changes and sex-stratified vascular differences across the adult lifespan. Furthermore, we evaluate VITAL on subjects with stroke-related neurological disorders, demonstrating its potential as a exploratory normative reference for anomaly analysis.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0382_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)

https://brain-development.org/ixi-dataset/

BibTex

@InProceedings{ZhaXia_Deep_MICCAI2026,
        author = { Zhang, Xiaoming AND Dannecker, Maik AND Cleary, Jon AND Falcetta, Daniele AND Yang, Kaiyuan AND Menze, Bjoern H. AND Shi, Yu AND Wegener, Susanne AND Ourselin, Sébastien AND Rueckert, Daniel AND Zuluaga, Maria A.},
        title = { { Deep Probabilistic Cerebrovascular Atlas } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16889},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes VITAL, a continuous implicit neural representation (INR) framework aimed at constructing a probabilistic cerebrovascular atlas. To address the multi-scale nature of the brain’s vascular tree and the spectral bias inherent in standard INRs, the authors integrate a dynamically routed Mixture-of-Experts (MoE) architecture combined with FiLM modulation. The method is evaluated on the healthy IXI dataset for atlas construction and on the TopCoW stroke dataset to demonstrate its potential for anomaly detection via test-time adaptation.

  • 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, Methodological Exploration: The application of INR to overcome the partial volume effects and aliasing inherent in discrete voxel grids for highly variable vascular structures is an interesting engineering exploration.

    2, Architectural Design: Introducing an MoE mechanism to theoretically decouple multi-scale spatial frequencies (e. g. , low-frequency global geometry of major arteries vs. high-frequency local details of distal branches) is a conceptually appealing approach to mitigating the spectral bias of standard MLPs (like SIREN).

  • 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.

    Summary: It appears the authors lack a practical understanding of real-world cerebrovascular challenges, which undermines the claimed clinical meaning of the work.

    1.Insufficient Performance for Clinical Cerebrovascular Applications The results have severe vascular discontinuity and alignment errors, which limits the framework’s clinical applicability. 1.1 Dataset Limitations vs. Classic Methods: The model relies on the IXI MRA dataset resampled to a $256 \times 256$ in-plane resolution. At this resolution and field strength, the primarily visible structures are the macro-vasculature (large to medium arteries like the CoW). For these macro-structures, standard discrete processing methods (e. g. , Frangi-based vessel tracking, graph-cut methods, or conventional 3D CNNs) already achieve reliable performance. The necessity of a complex MoE-INR architecture is not fully justified given that the dataset does not clearly capture the micro-vascular details that INRs are theoretically designed to preserve.

    1.2 Unclear Deviation Maps: The visual results presented in Figures 1, 2, and 3 are ambiguous, overly smoothed. Thus, it cannot provide sufficient evidence for reviewers to validate the method’s efficacy. Based solely on the visual performance in these figures, it appears the authors lack a practical understanding of real-world cerebrovascular challenges, which fundamentally undermines the claimed clinical significance of the work.

    This visual ambiguity is validated by the results in Table 1.An SSIM of 0.84 (reflecting an approximate 20% intensity reconstruction error), a Dice score of merely 0.71, and an Average Surface Distance (ASD) of 4.18 mm are highly problematic. Crucially, cerebrovascular imaging data is volumetrically dominated by macro-vasculature (large arteries). Because these large vessels constitute the vast majority of the target signal, baseline metrics are naturally inflated if the main arterial trunks are even partially captured. The fact that the proposed method still yields such subpar overall metrics. Despite the inherent statistical advantage provided by macro-vessel dominance, the method exposes severe algorithmic deficiencies. Consequently, the framework lacks the precision and reliability required for meaningful clinical tasks such as stroke localization or surgical planning. 1.3 Quantitative Discontinuities: The quantitative results in Table 1 show a whole-brain Connectivity Error of 183.69 ± 130.67, indicating notable vascular fragmentation. Furthermore, the framework yields a PSNR of 24.98 dB, an SSIM of 0.84, and an Average Surface Distance (ASD) of 4.18 mm. In cerebrovascular anatomy, where many medium vessels are 1-3 mm in diameter, an average error of 4.18 mm exceeds the size of the anatomical structures themselves, limiting the model’s ability to precisely quantify pathological deviations.

    2.Limitations in Modeling Discrete Topological Mutations The framework relies on MoE and FiLM modulation to capture anatomical variability. However, it is limited to handle discrete topological changes.

    • 2.1 On page 4, the modulation mechanism is described: “Following [10], the concatenated conditioning vector $\tilde{z}{i}=[z{i},c_{i}]$ is mapped to layer-specific scale ($\gamma$) and shift ($\beta$) parameters via a linear projection. “
    • 2.2 FiLM modulation applies global affine transformations (scaling and shifting) to the intermediate feature maps of the SIREN network. While modifying the frequency and phase of a continuous sine wave is effective for modeling global continuous morphing (e. g. , generalized widening or tortuosity due to aging), it is less suited for discrete local topological mutations. For example, the congenital absence of an A1 segment in the Circle of Willis is a binary (1-to-0) structural break. A global linear projection of a latent code cannot easily execute localized topological erasure without introducing artifacts (e. g. , semi-transparent vessels). The paper lacks theoretical or visual demonstration showing how FiLM and MoE accurately enforce these discrete topological graph changes. 3.Incomplete Baseline Comparisons The quantitative evaluation of the atlas reconstruction (Table 1) omits comparisons against standard clinical registration methods. 3.1 The Evidence: On page 5, the authors explain their baseline selection: “We compare VITAL against a state-of-the-art INR-based brain modeling baseline CiNeMA [9]. Traditional voxel-based registration methods are excluded as they do not employ a generative latent space for subject-specific reconstruction. “ 3.2 The Evidence: On page 5 (Preprocessing), the authors state: “For anatomical context, T1 images were segmented using FastSurfer [12] and rigidly registered to the MRA space via ANTs [4]. “ 3.3 The Critique: The authors use traditional registration tools (ANTs) for foundational data alignment but exclude state-of-the-art deformable registration methods (such as ANTs SyN) in their performance benchmarks. Since the primary metrics evaluated in Table 1 (e. g. , ASD and Connectivity Error) assess geometric alignment accuracy, the lack of a generative latent space in traditional methods does not justify excluding them from an alignment comparison. 4.Methodological Circularity in Clinical Anomaly Detection The experimental design for the TopCoW stroke dataset (Section 3) involves a circular dependency during the inference phase, which affects its clinical utility. 4.1 The Evidence: On page 7, the authors state: “Given the inherent imaging protocol variations between datasets and the severe morphological deviations in stroke subjects, relying solely on intensity reconstruction is prone to misalignment. Therefore, we employed supervised test-time adaptation using $\mathcal{L}{rec}+\mathcal{L}{seg}$ to investigate vascular anomalies, with the segmentation mask serving as a topological anchor. “ 4.2 The Critique: A primary objective of an anomaly detection system is to discover and segment abnormal vascular regions. By requiring the patient’s ground-truth segmentation mask ($\mathcal{L}_{seg}$) during the test-time adaptation phase to achieve alignment, the algorithm relies on the intended output to process the input. A model that requires a pre-existing mask of the anomaly to perform localization cannot be directly applied to a real-world diagnostic workflow. 5.Terminology Mismatch: “Spatio-Temporal Evolution” vs. Cross-Sectional Interpolation The manuscript uses longitudinal terminology to describe cross-sectional data manipulations. 5.1 The Claim: In the Abstract, the authors declare: “. . . we construct the first continuous spatio-temporal neural atlas of the human cerebrovascular system. . . “ and on page 7: “Fig. 4 presents the spatiotemporal evolution of VITAL atlas spanning the adult lifespan from 20 to 80 years. “

    5.2 The Reality: The dataset description on page 5 states: “In this study, we used MRA scans and corresponding T1 images from healthy adults (19-87 years old) in the IXI dataset [6]. “ The methodology relies on computing target latent codes via “Gaussian kernel regression over the trained population” (page 5).

    5.3 The Critique: The IXI dataset is strictly cross-sectional, containing single-timepoint scans of different individuals at various ages. Modeling true “spatio-temporal evolution” generally requires longitudinal data (repeated scans of the same subjects over time). Interpolating a latent code between subjects of different ages mathematically generates a cross-sectional age-related morphological mean, rather than a longitudinal aging trajectory. 6.Inconsistencies Regarding Discretization and Partial Volume Effects The motivation for utilizing an Implicit Neural Representation is to bypass the limitations of discrete grids; however, the preprocessing pipeline partially conflicts with this objective.

  • 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

    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?

    My recommendation for this manuscript is Reject.

    It appears the authors lack a practical understanding of real-world cerebrovascular challenges. While the integration of Implicit Neural Representations (INRs) and Mixture-of-Experts (MoE) for cerebrovascular modeling is an interesting engineering exploration, the paper suffers from several flaws across its clinical logic, experimental design, and conceptual claims.

    The major factors driving this decision are:

    1.Methodological Circularity in Clinical Evaluation: The most fatal flaw is the logical tautology in the anomaly detection experiment (Section 3). By requiring the ground-truth segmentation mask of the stroke anomaly during the test-time adaptation phase to achieve anatomical alignment, the framework relies on the intended diagnostic output to process the input. This circular dependency completely invalidates the model’s utility in any realistic clinical triage workflow.

    2.Insufficient Empirical Performance and Unfair Baselines: The quantitative results demonstrate severe topological fragmentation and alignment errors (e.g., an Average Surface Distance of 4.18 mm and a Connectivity Error of 183.69). For cerebrovascular anatomy, an error margin of >4 mm often exceeds the diameter of the vessels themselves, rendering the atlas clinically unreliable. Furthermore, the authors deliberately excluded standard, state-of-the-art deformable registration methods from their baseline comparisons, making it impossible to objectively assess their claims of superior topological alignment.

    3.Mathematical and Conceptual Inadequacies: The framework’s reliance on global affine transformations (FiLM) is mathematically ill-equipped to model discrete, localized topological mutations (e.g., congenital vessel absences), leading to ambiguous probabilistic outputs rather than deterministic structures. Additionally, the manuscript makes severe scientific overclaims by repeatedly characterizing the latent interpolation of static, cross-sectional data (the IXI dataset) as longitudinal “spatio-temporal evolution.”

    Thus, the combination of a flawed clinical validation paradigm, subpar quantitative performance, and significant conceptual misrepresentations indicates that this work requires profound restructuring in both its mathematical priors and experimental design. It does not meet the scientific and clinical standards expected for MICCAI.

  • 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.

    1.Regarding “Circularity in TopCoW analysis” The authors argue that the mask used for test-time adaptation (TTA) “contains only observed vessel… labels,” “doesn’t supervise the deviation map,” and merely “anchors alignment.” This defense reveals a critical misunderstanding of cerebrovascular pathology (specifically in stroke datasets like TopCoW) and the mechanics of image registration/alignment.

    The Logical Flaw: In clinical, a stroke patient with occlusion, a major vessel (e.g., the MCA) is missing from the scan.

    If the VITAL framework uses this topologically incomplete pathological mask as the explicit target to optimize the latent code ($Z_t$) and spatial alignment parameters of a healthy atlas, the optimization process will inevitably be pathologically biased.

    The network will forcefully warp the healthy atlas to minimize the loss against an empty space where the vessel should be.

    Conclusion: You cannot align a healthy topological prior to a pathologically broken target by using the broken target’s mask as the primary geometric anchor.

    This will distort the latent space alignment, making the subsequent deviation map a meaningless product of misregistration rather than a true reflection of the anomaly.

    2.Regarding “Metric interpretation” The authors argue that high Dice indicates “memorization” and that the metrics should be read merely as an ablation showing MoE is better than CiNeMA. They state VITAL is a “normative reference.”

    The Physical Flaw: While a probabilistic atlas should not memorize individual topology, it must provide a baseline accurate enough to be clinically meaningful.

    The authors ignore the absolute scale of their error. According to Table 1, the Average Surface Distance (ASD) is 4.18 mm. In human cerebrovascular anatomy, the diameter of major arteries (e.g., ACA, MCA, PCA) is typically between 1.5 mm and 3.5 mm.

    Conclusion: If the reference has an intrinsic geometric uncertainty/error larger far than the physical diameter of the vessels, it cannot reliably compute a deviation map for anomalies (such as a 2mm occlusion or a 3mm aneurysm).

    The margin of error dwarfs the physiological signal and the results is meaningless for clinical.

    3.Regarding “Topology Mutation” The authors state that congenital absences appear as “regions of low vessel probability and high uncertainty rather than hard topology deletions.”

    Acknowledgment: This is a fair mathematical admission.

    Limitation: However, it confirms my initial critique: the framework blurs across distinct discrete topological states (e.g., presence vs. absence of the A1 segment) by outputting a muddy average probability.

    4.Regarding “Age-conditioned vs. longitudinal terminology”

    I acknowledge and appreciate the authors’ concession to revise the terminology from “spatio-temporal evolution” to “cross-sectional lifespan morphological modeling.”

    This resolves the scientific overclaiming issue regarding the dataset’s nature.



Review #2

  • Please describe the contribution of the paper

    the main contribution consists of a cerebrovascular atlas obtained by training a mixture of experts model to learn details of cerebrovasculature at different scales

  • 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 work is thorough and well explained/organized with clear applications and strong well presented results

  • 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.

    -I found the notion of “population-level conditioning” introduced in the paper somewhat unclear. From my understanding, the conditioning appears to be performed at the patient level. This distinction should be clarified, as it is important for interpreting the proposed method.

    -Although the challenge of modeling vasculature across different scales is well recognized, the paper does not provide sufficient ablation studies to demonstrate how the MoE architecture effectively addresses this issue. In particular, it is unclear how the number of experts is selected. Figure 2c provides some qualitative insight, but it is based on a single sample and does not constitute a comprehensive evaluation.

  • 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

    In the paragraph that begins with “In this work” on the second page I think “patterns of vesicular tree” is a typo

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a well-motivated and technically sound approach to modeling vascular structures, addressing a relevant and challenging problem. The methodology is clearly described and generally well organized, making the contributions easy to follow. In addition, the experimental results are strong and convincingly demonstrate the effectiveness of the proposed approach.

    The work is thorough and shows careful design choices, with clear practical applications. Overall, despite some minor issues in clarity and additional analysis that could further strengthen the paper, I believe the contribution is solid and meets the standards for acceptance.

  • 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 believe there is real value in the work, and I thank authors for addressing reviewers concerns



Review #3

  • Please describe the contribution of the paper

    The authors present the first spatiotemporal atlas of the human cerebrovascular system constructed using implicit neural representation with an MoE architecture.

  • 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 is a very well-written and comprehensive piece of work.

    The main novelty lies in the application. The overall framework based on which this method is developed, i.e., CINA/CINeMA, has already been around for a while and is popular for atlas construction but has never been applied to cerebrovascular atlas construction.

    In terms of the framework, the authors further incorporate an MoE architecture beyond the original CINeMA, and demonstrate its effectiveness in improving the overall performance of the proposed method. The motivation for the MoE architecture is well-explained. For an 8-page short paper, the authors provide a lot of experimental results to support the usefulness of the MoE architecture (Table 1), characterization of how the proposed MoE architecture work indeed result in different experts specializing in different areas of the vasculature (Fig. 2(c)), and usefulness for downstream tasks (clear separation of Normative vs. Pathological in Fig. 3(c)).

    Key design choices are well-justified. For example, the authors employ a label-stratified coordinate sampling approach to ensure sufficient sampling of the foreground coordinates, which are relatively sparse.

  • 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.

    Although the application of INR to cerebrovascular atlas construction is quite new and, in fact, first of its kind, the novelty of the method itself is somewhat limited. From a method stand point, the authors are aware and acknowledge previous works, e.g., CINeMA [9] and MoE for INR [5], which greatly influenced this work. The proposed method does not introduce much new stuff beyond a combination of [9] and [5].

    There are also some ambiguities in the description of the method. 1.In Section 2.1, what is vessel intensity value I(x) referring to? Is it the intensity values of the MRA images? 2.Also in section 2.1, the author claim that affine transformation is used. However, the formulation of the transformation is rotation plus translation makes it sounds like a rigid transformation is used? It would be helpful to clarify. 3.Fig. 2(b) and the discussion thereof are hard to follow. There are a lot of information in Fig. 2(b) and I believe it would be helpful to provide more explanation. Specifically, it would be helpful to mention that, e.g., VITAL tend to assign larger probability than competing methods, which demonstrates that less smoothing occurs spatially.

  • 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

    There are some places where additional discussions/explanations would be appreciated. 1.On page 2, the authors mention that “… SIREN … difficult to simultaneously encode the low-frequency … and the high-frequency …”. Isn’t the sinusoidal activation patterns of SIREN designed specifically to mitigate this very issue?

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Despite the limited novelty, I believe the work is still of interest to the MICCAI community for a couple of reasons. 1.The application of INR to modeling vasculature of the brain is novel. 2.Having a spatiotemporal cerebrovascular atlas is not only cool in its own right but of great importance to research and clinical applications.

  • 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 clarifications provided by the authors address the ambiguities. Although the novelty is somewhat limited, I still believe this work is highly relevant to the MICCAI community. In particular, using INR for the task of cerebrovascular atlas construction is not only interesting but also relevant for clinical applications.

    Authors’ responses to R1 is very informative and highlights some of R1’s misunderstand of the scope and clarifies the intended use case of this work. The responses are well-written and convincing.



Author Feedback

We sincerely thank all experts for their constructive comments [R1,R2,R4]. We are encouraged that R2 and R4 found our work well-motivated, technically sound, and of clear interest to MICCAI. We clarify all key concerns below.

Scope [R1]. Several concerns appear to arise from interpreting VITAL as a segmentation, diagnostic, or surgical-planning system. This is not our intended claim. VITAL is a continuous probabilistic population atlas and normative reference: it models healthy population variability and enables atlas-to-subject comparison, not a standalone diagnostic, segmentation, or navigation tool.

Circularity in TopCoW analysis [R1]. This concern follows from the same scope misinterpretation. The masks used in test-time adaptation contain only observed vessel and brain-region labels, which can be obtained from existing segmentation tools. They don’t contain anomaly labels, lesion masks, stroke labels, or ground truth about pathology. The deviation map is computed after atlas projection as the mismatch between the healthy population prior and observed anatomy (Figs. 1 and 3). So the mask anchors alignment; it doesn’t supervise the deviation map. We will clarify this more.

Metric interpretation [R1]. The metrics measure atlas-to-subject reconstruction quality of a probabilistic population prior, not supervised segmentation accuracy. Extremely high Dice might indicate memorization of individual topology rather than a meaningful population prior, the opposite of what a normative atlas should achieve. Table 1 should be read mainly as an ablation: MoE (K=4)+Lmoe improves over CiNeMA across most metrics, including PSNR (+3.2 dB), SSIM (+0.05), ASD (-0.98 mm whole-brain, -0.52 mm CoW), conn. error (-112), and center. distance (-0.95 mm). We will revise the wording to avoid implying surgical-grade precision, which is not claimed.

Baseline comparisons [R1]. Table 1 evaluates conditional reconstruction within generative atlas frameworks; CiNeMA is the closest comparable baseline. Fig. 2 includes NLR/MultiMorph to illustrate population-level wash-out in discrete atlases, not to claim superiority over deformable registration in a geometric alignment benchmark. ANTs was used solely for intra-subject (T1 to MRA) rigid preprocessing to align anatomy labels, which is a standard procedure, and not as an atlasing baseline.

Age-conditioned vs. longitudinal terminology [R1]. We agree that IXI is cross-sectional and that spatio-temporal evolution overstates the design. We will use age-conditioned atlas and cross-sectional lifespan morphological modeling instead. The intended contribution, a continuous implicit function conditioned on age modeling population-level morphological variation across the adult lifespan, is a standard approach in normative brain atlasing.

Topology Mutation [R1]. FiLM alone does not guarantee deterministic binary graph edits, but VITAL uses spatial latent codes rather than a single global latent code for local variation without global compensation, with MoE partitioning cross-scale capacity. Anatomical variants are represented probabilistically: congenital absences appear as regions of low vessel probability and high uncertainty rather than hard topology deletions.

Clarifications [R2,R4]. 1) VITAL conditions at two levels: z encodes subject-specific variation; c encodes population-level covariates. 2) K=4 was selected as a capacity-cost trade-off supported by the ablation in Table 1 (K=2 vs. K=4); Fig. 2c is qualitative and representative, not a comprehensive evaluation. 3) Yes, it should be ‘vascular’. 4) I(x) denotes MRA intensity. The transformation is rigid to avoid over-warping sparse vascular branch. 5) Fig. 2b shows VITAL assigns more mass to high vessel probability than voxel-based atlas methods, indicating less spatial wash-out. 6) SIREN mitigates spectral bias, but a monolithic SIREN shares one parameter set across scales, causing capacity competition; MoE explicitly partitions this capacity (Fig. 2c).




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 paper proposes VITAL, an implicit neural atlas with a Mixture-of-Experts architecture for cerebrovascular modeling, applied to anomaly analysis on stroke subjects. Several substantive concerns are raised across the reviews. The methodological novelty appears limited as the approach largely combines CINeMA and MoE-for-INR, and warrants stronger articulation. More fundamentally, the anomaly analysis exhibits a circular dependency where the test-time adaptation relies on the patient’s segmentation mask through L_seg, and the validity of the resulting deviation maps as a clinical signal needs to be defended. The reported metrics, including an average surface distance exceeding the diameter of medium cerebral vessels, are also difficult to contextualize given the omission of standard deformable registration baselines. Further concerns include the characterization of cross-sectional latent interpolation as “spatio-temporal evolution,” and ambiguities in population-level versus subject-level conditioning and in the transformation formulation. The rebuttal should address the major reviewers’ concerns including these points.

  • 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.

    This paper proposes VITAL, an implicit neural atlas with a Mixture-of-Experts architecture for continuous probabilistic modeling of the cerebrovascular system, applied to atlas-to-subject deviation analysis on stroke data. The reviewers acknowledge that applying INR to cerebrovascular atlas construction is the first of its kind and clinically relevant, and that the MoE design is well motivated and supported by ablations. The rebuttal clarified the intended scope as a probabilistic population atlas and normative reference rather than a standalone diagnostic system; under this scope, the masks in test-time adaptation function as alignment anchors and the metrics are appropriately read as an ablation against the closest generative atlas baseline. The authors are encouraged to revise the “spatio-temporal evolution” terminology and to soften wording suggesting surgical-grade precision, as committed. The novelty lies primarily in adapting existing components, but the application is genuinely new for this anatomy and the empirical support is convincing.



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.

    Reviewers summarized strengths of your work, novel aspects and noted that your solution to the specific application area may be of potential interest. The rebuttal helped to clarify open issues and misunderstandings. Would the paper be accepted, authors are strongly encouraged to improve the final write-up by including recommendations by reviewers, in particular to improve clarity in order to avoid misunderstandings.



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.

    Based on the reviewers’ evaluations, the authors’ rebuttal, and the initial meta-review, the paper is recommended for acceptance. While Reviewer #1 raised serious concerns about clinical utility, insufficient performance, and terminology, two other expert reviewers (R2 and R4) found the work well-motivated, technically sound, and of clear interest to the MICCAI community. The rebuttal convincingly clarified the scope as a probabilistic population atlas and normative reference, not a standalone diagnostic or segmentation tool, and addressed the major points by revising the “spatio-temporal” terminology to cross-sectional age-conditioned modeling, explaining the role of the mask in test-time adaptation, and contextualizing the metrics as an ablation of the MoE architecture relative to the closest INR baseline. The accepting reviewers’ post-rebuttal assessments reaffirmed their support, noting that the responses resolved ambiguities and underscored the novelty and relevance of applying implicit neural representations to cerebrovascular atlas construction.



back to top