List of Papers Browse by Subject Areas Author List
Abstract
Cerebral small vessel disease (CSVD) markers, specifically enlarged perivascular spaces (EPVS) and lacunae, present a unique challenge in medical image analysis due to their radiological mimicry. Standard segmentation networks struggle with feature interference and extreme class imbalance when handling these divergent targets simultaneously. To address these issues, we propose a morphology-decoupled framework where Zero-Initialized Gated Cross-Task Attention exploits dense EPVS context to guide sparse lacune detection. Furthermore, biological and topological consistency are enforced via a mixed-supervision
strategy integrating Mutual Exclusion and Centerline Dice losses. Finally, we introduce an Anatomically-Informed Inference Calibration mechanism to dynamically suppress false positives based on tissue semantics. Extensive 5-folds cross-validation on the VALDO 2021 dataset (N = 40) demonstrates state-of-the-art performance, notably surpassing task winners in lacunae detection precision (71.1%, p = 0.01) and F1-score (62.6%, p = 0.03). Furthermore, evaluation on the external EPAD cohort (N = 1762) confirms the model’s robustness for large-scale population studies. Code: https://github.com/lucas-ucl/Lacune-EPVS-MTL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5165_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/lucas-ucl/Lacune-EPVS-MTL
Link to the Dataset(s)
VALDO dataset: https://zenodo.org/records/4520773
EPAD dataset: https://ep-ad.org
BibTex
@InProceedings{HeLuc_AUnified_MICCAI2026,
author = { He, Lucas AND Li, Krinos AND Zhang, Hanyuan AND He, Runlong AND Ingala, Silvia AND Lorenzini, Luigi AND de Bruijne, Marleen AND Barkhof, Frederik AND Davies, Rhodri AND Sudre, Carole H.},
title = { { A Unified Framework for Joint Detection of Lacunes and Enlarged Perivascular Spaces } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of the paper lies in the proposed method itself (Section 2.1), i. e. the dual-decoder network with a gated cross-task attention mechanism designed in a way that it separates tubular (PVS) from spherical (lacunes - they are not always spherical btw) structures.
- 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 topic is clearly of interest, and reliable tools for joint lesion detection, segmentation, and quantification are very much needed. The study therefore addresses a real methodological gap.
- The model is evaluated against manually annotated data, which remains the gold standard for lesion detection and segmentation.
- 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.
- Section 2.1, which is in my view the most interesting and original part of the work, is treated too superficially. Several of the claims made in this section would benefit from more thorough discussion and development.
- Section 2.2 closely resembles a standard masking procedure, and it is therefore unclear to me the extent to which it constitutes a methodological contribution in its own.
- Given that the proposal focuses on the detection of lacunes and PVS, it is unclear how applying the model to EPAD constitutes a rigorous evaluation setting. A correlation between visual scores and predicted values alone does not demonstrate that the method performs accurately.
- Generalisation to external datasets was not assessed. At this stage, it is not sufficient to demonstrate that deep learning models perform well when training and test data are similar, as this is already well established. The more relevant question is whether the framework remains reliable under meaningful domain shift, for example, when applied to anisotropic imaging data.
- The authors emphasise numerically best results (numerically best in terms of what? mean? dispersion?) without formally testing for paired differences, making it unclear whether the proposed method meaningfully improves over either competing approaches or baseline models (Tables 1–3).
- Other new methods that did not participate in the VALDO challenge addressing similar tasks were not considered in the evaluation, including SHiVAi (https://github. com/pboutinaud/SHiVAi) and a recent dedicated PVS segmentation framework published here https://doi. org/10.3389/fneur. 2025.1540923.More broadly, the field of PVS segmentation has advanced substantially in recent years, and the benchmark would benefit from including a more comprehensive and up-to-date set of competing methods.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
See major weaknesses
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes a unified multi-task deep learning framework for the joint detection of lacunes and enlarged perivascular spaces (EPVS), two key biomarkers of cerebral small vessel disease that are difficult to distinguish due to similar imaging appearance but distinct morphology. The method introduces a morphology-decoupled architecture with a shared encoder and dual decoders, coupled via a unidirectional gated cross-task attention mechanism that leverages dense EPVS context to guide sparse lacune detection. The framework further incorporates mixed supervision (combining dense masks and weak labels), topology-aware losses (e.g., centerline Dice), and a mutual exclusion constraint to enforce biological consistency. Additionally, an anatomically-informed inference calibration module suppresses false positives in implausible regions using tissue priors. The approach is evaluated on the VALDO 2021 dataset and externally validated on the EPAD cohort, demonstrating improved performance over prior methods, particularly in lacune detection precision and F1-score.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper addresses a clinically relevant and technically challenging problem, namely the joint detection of radiologically similar but morphologically distinct CSVD markers under severe class imbalance. A key strength is the integration of domain knowledge into the modeling pipeline, including morphology-aware decoupling, anatomical priors, and biologically motivated constraints such as mutual exclusion and topology preservation. The proposed cross-task attention mechanism is well-motivated and provides a reasonable way to exploit contextual relationships between EPVS and lacunes without feature entanglement. The use of mixed supervision is practical and aligns with real-world annotation scarcity. Empirically, the method demonstrates consistent improvements over strong baselines and prior challenge winners, particularly in reducing false positives and improving precision, which is critical for clinical usability. The inclusion of external validation on a large cohort (EPAD) further strengthens the claims of generalizability and practical applicability.
- 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 cross-task attention design is relatively simple, and its advantage over standard attention or feature fusion strategies is not deeply analyzed.
2.The anatomical calibration step relies on external segmentation (FastSurfer), introducing additional preprocessing complexity and potential error propagation.
3, Evaluation on the main dataset (VALDO, N=40) is limited in size, raising concerns about robustness and statistical power despite cross-validation.
4.The external validation on EPAD uses weak labels only, making it difficult to fully assess voxel-level segmentation accuracy and limiting the strength of generalization claims.
5.The method prioritizes precision over recall, and while this reduces false positives, it may lead to missed lesions, which is not fully explored from a clinical perspective.
6.Some claims regarding “state-of-the-art” and clinical readiness are slightly overstated given the dataset size and evaluation setup.
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
the paper offers a well-engineered and clinically meaningful contribution, but with moderate novelty and some limitations in validation
- 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.
the paper offers a well-engineered and clinically meaningful contribution, but with moderate novelty and some limitations in validation
Review #3
- Please describe the contribution of the paper
5.This article describes the provision of a new framework for jointly detecting two different markers of Cerebral Small Vessel Disease (CSVD): 1) Enlarged Perivascular Spaces (EPVS) and 2) Lacunes. This Contribution will provide: Architectural Innovation , A dual-decoder network with a Zero-Initialized Gated Cross-Task Attention mechanism that uses the high-density spatial context of an EPVS to assist in the detection of a very low-density Lacune while maintaining the lack of feature entanglement; Loss Function Design , Incorporation of Mutual Exclusion and Centerline Dice Losses into the loss function to maintain biological and topological consistency between the two markers; Refinement of Inference , Development of an Anatomically driven inference calibration process that incorporates the loss of a visually destructive object based on distance fields around body tissue to change the binarization threshold and reduce the number of false-positive results identified within the exclusion zone.
- 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.
6.The manuscript introduces a novel morphology-decoupled framework utilizing a sophisticated multi-task architecture to address the significant challenge of radiological mimicry between enlarged perivascular spaces (EPVS) and lacunes. This architecture incorporates a pivotal Gated Cross-Task Attention Module, enabling a unidirectional information flow to leverage the dense spatial context of the perivascular anatomy to guide the sparse detection of lacunes, thereby preventing feature entanglement arising from their physiological correlation. The effectiveness of this architecture is further bolstered by two explicit biological and topological constraints—namely, mutual exclusion and soft-centerline Dice losses—which represent a shift toward biology-informed deep learning optimization as opposed to purely data-driven approaches. The implementation of Anatomically-Informed Inference Calibration using distance fields is particularly significant, as it suppresses false positives in biologically implausible regions, thus enhancing specificity for lacune detection. These architectural and optimization refinements yield statistically significant improvements in precision (71.1%, p=0.01) and F1-score (62.6%, p=0.03) for lacune detection compared to the state-of-the-art task winners on the VALDO 2021 dataset. Finally, the framework exhibits robust generalizability through its application to the large-scale external EPAD cohort (N=1762), underscoring its utility for population-level vascular burden quantification.
- 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.
7.The methodology used in this study supports their results. However, the framework has several observable limitations. The anatomical calibration reduces false positives but also creates a trade-off in sensitivity, especially in detecting smaller lesions. The hyperparameter settings employed lack a systematic sensitivity test and, therefore, will not be applicable to a variety of different studies. Although a decoupled strategy is used in detecting EPVS, the task branch feature interference will continue to be an ongoing issue because the F1-score for EPVS detection was only 53.7%. In addition, since weak labels were relied upon to validate the EPAD cohort, the absolute voxel-wise precision could not be assessed at greater detail across multiple types of clinical practice.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
12.In conclusion, the manuscript presents a technically rigorous solution to a challenging medical imaging problem. The authors are encouraged to utilize the rebuttal phase to specifically address the identified sensitivity trade-offs and offer further justification for the hyperparameter selection process. Clarifying these critical aspects would substantiate the framework’s practical utility and strengthen its overall contribution to the field.
- 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?
14.The suggestion for a weak accept comes from the methodology used in the paper; in the way that the author addresses CSVD radiological mimicry by using a morphology-decoupled approach, and the addition of cross-task gating attention and anatomical prior integration using distance-filed calibration have shown an improvement in statistically significant performance, especially for lacune precision detection. However, the score is offset by concerns with respect to trade-offs associated with the trait of anatomical constraints and lack of proper sensitivity analysis for the empirical hyperparameter settings. A proper rebuttal of these aspects would likely give greater justification for an available acceptance.
- 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 #4
- Please describe the contribution of the paper
This paper proposes a morphology-decoupled multi-task framework for joint detection of enlarged perivascular spaces (EPVS) and lacunae in cerebral small vessel disease. The method introduces a gated cross-task attention mechanism to leverage EPVS context for lacune detection, along with mixed supervision (mutual exclusion and centerline Dice losses) and an anatomically-informed inference calibration strategy. The approach is evaluated on the VALDO 2021 dataset and externally on the EPAD cohort.
- 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.This is a clinically relevant problem. Joint modeling of EPVS and lacunae is important and challenging due to radiological similarity and class imbalance.
2.The design motivation has been done in a thoughtful manner. The idea of morphology decoupling and leveraging dense EPVS context to guide sparse lacune detection is intuitive and well-motivated.
3.Comprehensive evaluation protocol: The use of multiple metrics, statistical testing, and external validation on a large cohort (EPAD) is commendable.
- 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 empirical gains are limited and inconsistent. While improvements are reported in precision and F1-score, the Dice scores (38% for EPVS and 42% for lacunes) are relatively low and even inferior to the VALDO winner, which weakens the claim of state-of-the-art performance. Given that Dice and surface-based metrics are standard for segmentation quality, this discrepancy raises concerns about the overall effectiveness of the method.
2.The methodological novelty is moderate, as the proposed components—cross-task attention, topology-aware losses, and post-hoc calibration—are incremental extensions of existing ideas. More important information would have been anatomical cues (including higher weightage for basal ganglia region where EPVS are shown clinically to occur more). Their integration is reasonable but does not constitute a fundamentally new framework. Moreover, the ablation improvements are marginal, making it unclear whether each component contributes meaningfully.
3.The external validation on EPAD, while valuable, relies on weak labels (presence and visual scores), limiting the strength of conclusions regarding voxel-level segmentation performance and generalization. Additionally, the framework introduces multiple components (attention, mixed supervision, calibration), but no analysis of computational complexity or inference cost is provided, which is important for large-scale deployment.
4.Finally, although statistical significance is reported, the practical significance of improvements is unclear, especially given the relatively low absolute performance and mixed metric outcomes.
5.The comparison lacks several important lacune-specific detection methods, including recent foundation-model-based approaches (e.g., SAM-based lacune detection: Deepika et al, 2025), contrastive learning frameworks, and dedicated detection–segmentation pipelines such as MixLacune (Luther et al, 2021). Additionally, prior-informed and multi-stage detection models (al-masni et al, 2020) widely used in the VALDO challenge ecosystem are not considered. As a result, the evaluation does not fully reflect the current state of the art in lacune detection, particularly for distinguishing lacunes from radiological mimics such as EPVS.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
Please see the weaknesses section above. To be competitive for MICCAI, the work would benefit from clearer and consistent improvements across all key metrics (especially Dice), stronger justification of each component’s contribution deeper analysis of failure cases and metric discrepancies evaluation with stronger voxel-level ground truth in external datasets and extensive state-of-the-art validation.
- 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.
I can change my decision to weak accept since the comments have been addressed in sufficient detail. Though, the novelty could have been better addressed.
Review #5
- Please describe the contribution of the paper
Detection of lacunes and enlarged perivascular spaces (EPVS) is a challenging task in the medical domain. To address these and feature-interference and class-imbalance problems faced during deep learning model training, the authors proposed a multi-task morphology-decoupled framework in which Zero-Initialised Gated Cross-Task Attention exploits dense EPVS context to guide sparse lacune detection. In the proposed method, the authors also introduced biological and topological consistency and an Anatomically-Informed Inference Calibration mechanism to dynamically suppress false positives based on tissue semantics. Evaluations are conducted on the VALDO 2021 dataset, and the proposed method achieved superior scores.
- 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 is well written and explores a challenging research area. The proposed morphology-decoupled framework is interesting and novel. The proposed method has achieved superior performance compared to related work.
- 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 manuscript would benefit from a clearer explanation of how the proposed model components address challenges such as feature interference and class imbalance. While these aspects are mentioned, they are not sufficiently analysed or supported by dedicated empirical evaluation, particularly in the case of feature interference. The use of the Tversky loss for handling class imbalance is very common in the literature. An ablation study investigating the choice of the weighting parameter (\lambda_excl) is currently missing and would strengthen the methodological justification. Given the number of components in the framework, an analysis of computational complexity or efficiency would also be valuable to assess its practical applicability. The experimental comparisons could be expanded to include additional more relevant methods (e.g., PINGU) to provide a more comprehensive evaluation.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents an interesting formulation and addresses a relevant problem, with a reasonably designed framework and promising results. While the approach appears sound and shows potential, the experimental validation could be strengthened and some methodological aspects, such as handling of feature interference and computational complexity, would also benefit from clearer justification.
- 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.
I thank the authors for their detailed rebuttal. The response clarifies several of my concerns, particularly regarding the intended scope of the method. Overall, the rebuttal addresses several of my concerns and improves confidence in the contribution. The paper presents a clinically relevant problem, a well-motivated framework, and promising results, including external validation on EPAD. Although some methodological and efficiency analyses could be strengthened, I consider the remaining limitations addressable in revision.
I would raise my score to Accept.
Author Feedback
We thank all reviewers for noting the clinical relevance of the problem, the novelty of our morphology-decoupled framework (R3, R5), our integration of anatomical priors and biological constraints (R2, R3), the reduction in FPs and significant gains in precision (R2, R3), and our external validation on EPAD (R2, R3, R4). We address remaining concerns thematically.
Missing comparators (R1, R4, R5): Our work targets joint detection of EPVS and lacunes under radiological mimicry; the cited references address adjacent but distinct problems. Zhao et al. (Front Neurol 2025) use four parallel single-task decoders without cross-task interaction, mutual exclusion, or mimicry handling, on a single-institution cohort (N=105 + N=58 same-hospital “external”, single ethnicity) with correlation-based metrics and no detection F1.MixLacune participated in VALDO 2021 (F1=6.25; Sudre 2024); the Task-3 winner we statistically surpass already substantially outperformed it. Al-Masni 2020 is a two-stage cerebral microbleed detector; it predates VALDO and does not address lacunes or EPVS. SHiVAi is PVS-only; the 2D SAM-based method (Deepika 2025) targets lacunes only, while we already benchmark against VISTA-3D, a 3D SAM-family foundation model. These references can be cited in revision.
EPAD validation (R1-R4): Our two-dataset design deliberately separates voxel-level methodological evaluation (VALDO, Table 1: DSC, NSD, F1) from population-scale clinical validation (EPAD, N=1762), the standard pipeline in CSVD research when voxel-level annotation is infeasible at scale. EPAD spans multiple countries, scanner, and acquisition protocols, providing the meaningful domain shift R1 requests. Beyond correlations, we report a Balanced Accuracy of 64.9±2.1% for lacune presence as a non-correlation classification result.
Novelty (R1, R2, R4, R5): Our novelty operates at the system level: mimicry-aware joint modelling of EPVS and lacunes, beyond prior multiclass approaches. We respectfully note R1 may have read §2.2 as binary masking; Eq. 2 defines a continuous spatially adaptive threshold via a truncated distance field with tanh modulation, not a hard mask. The Zero-Initialised GCTA (§2.1) differs from standard attention: zero-initialising V (ReZero) ensures identity at initialisation, enabling stable joint training, while standard attention entangles tubular and ovoid features.
Evaluation Metrics (R3, R4): Dice is not the appropriate metric for objects of 5 to 20 voxels, and our results in fact approach the human inter-rater ceiling for this task. VALDO inter-rater statistics (Sudre 2024) report F1 = 19.6% (EPVS), 48 to 56% (lacune); Dice = 52.6 (EPVS), 59.0 (lacune). Our F1 (53.7/62.6%) matches or exceeds inter-rater agreement; our Dice (38.1/42.4%) approaches the human upper bound. The Tversky loss (α=0.1, β=0.9) explicitly optimises recall, intentionally misaligning Dice and the training objective. Detection F1 is the clinically operative metric.
Precision-recall trade-off (R2-R5): Our framework deliberately favours precision over recall, which is clinically appropriate at population scale: FPs inflate prevalence, bias dose-response in cohort studies, and impose prohibitive review burden, as Sudre 2024 notes for FP-heavy lacune baselines. The ablation supports this: combined losses and calibration reduce lacune FP from 3.9 to 0.7 per subject, an 82% reduction. Missed cases are predominantly small or borderline lesions where ground truth is itself rater disputed.
Calibration and hyperparameters (R2-R5): Anatomical priors are standard in CSVD pipelines; our calibration is segmentor-agnostic and low-cost. Real-time inference is not required. R4’s basal-ganglia weighting is implicit, as deep grey matter falls within Zone-1.We did not perform a systematic λ/γ/λ_excl sweep; λ=γ=0.5 was a neutral midpoint; compute statistics will be reported where space permits.
We will also soften ‘SOTA’ and ‘clinical readiness’ wording (R2).
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.
This paper proposes a unified framework for the joint detection of lacunes and enlarged perivascular spaces (EPVS). Most reviewers support weak accept, recognizing its clinical relevance and biologically informed design, while raising concerns on novelty, validation, and metric performance. I opt for Invite for Rebuttal such that the authors can clarify concerns raised by the reviewers.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
As Reviewers point out, the paper offers a well-engineered and clinically important contribution, although with moderate novelty and some limitations in the validation. However, given the difficulties in limited datasets with annotations, and the recognition of these imaging features (e.g. a big PVS can be classed as lacune and viceversa), it is fair to acknowledge that validation in the multicentre european-wide dataset mentioned (EPAD) gives confidence in the value and applicability of the work.
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.
This paper received 5 reviews. After rebuttal, 3 reviewers give positive score and 2 give negative score. After reading the paper, comments, and rebuttal, I lean to acceptance of this work. Although the methodological novelty is moderate, it is overall a well-motivated framework with promising results, including external validation.
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.
I would like to start echoing a major strength highlighted by the reviewers. This is an clinically relevant and interesting problem to solve, as the authors clearly underscore, these two type of abnormalities might have a relationship and segmenting them together would be beneficial. However, there were some concerns in the original revision that were raised by the reviewers.
In terms of the rebuttal, I have mixed feelings and I think at times, the authors are fairly evasive and I do not fully agree with all the answers given. Having said that, my disagreements and issues with the rebuttal are minor and do not eclipse the comments were I actually agree (e.g. to an extent the missing comparators, the EPAD validation and to an extent the evaluation metrics). With that in mind, and considering that 2 of the reviewers were happy with the rebuttal and one of them already suggested weak accept without changes during rebuttal, I would consider that the positives slightly outweigh the negatives.