List of Papers Browse by Subject Areas Author List
Abstract
Intracranial aneurysms (IAs) carry high mortality upon rupture, and their accurate segmentation from CTA is essential for clinical intervention. However, many IAs are small and morphologically resemble normal vascular bends, rendering them difficult to distinguish by intensity alone. Voxel-only networks frequently misclassify these subtle protrusions as tortuous healthy vessels, while point-based architectures capture geometry but lack dense semantic context. To bridge this gap, we propose GeoP2VNet, a geometry-aware method that injects explicit surface geometry into a voxel backbone via three modules: Geometric Feature Extraction (GFE) computes rotation-invariant point descriptors; a differentiable Point-to-Voxel (P2V) Gaussian splatting layer projects sparse geometric cues onto aligned voxel features; and Dual-Gated Residual Fusion (DGRF) selectively amplifies geometric priors at anatomically significant regions while suppressing them in normal segments. On 205 CTA cases (266 lesions) with 5-fold cross-validation, GeoP2VNet achieves a Dice of 0.770, outperforming nnU-Net (0.663) and SwinUNETR (0.578), while reducing false positives by 92%, from 0.317 to 0.024 per case. Code is available at https://github.com/somtiannes/GeoP2VNet.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0687_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/somtiannes/GeoP2VNet
Link to the Dataset(s)
N/A
BibTex
@InProceedings{QinTia_GeometryAware_MICCAI2026,
author = { Qin, Tian AND Zhang, Zhongji AND Feng, Yuqi AND Huang, Tianyu AND Yin, Zihan AND Luo, Tao},
title = { { Geometry-Aware Point-to-Voxel Fusion via Gaussian Splatting for Intracranial Aneurysm Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose extracting explicit geometric features from the vessel surface and combining them with intensity- or voxel-based features. This is done via three modules: a geometric feature extraction (GFE) module to extract geometric features, a Gaussian splatting layer to project these features into voxel features, and a dual-gated residual fusion to combine both feature types for intracranial aneurysm segmentation.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Combining geometric and intensity features will address the limitations of each in aneurysm segmentation. Also, the fact that the geometric features are not learned may provide an interpretable way to assess their impact on the model.
- 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 ablation may have provided a way to assess geometric features in more detail, for instance, by making some changes or alterations to the cloud points.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The term CTA (Computed Tomography Angiography) is not defined in the paper. The method for reproducing the geometric feature is somewhat unclear, while a handcrafted guide on how to design such features would improve the paper.
- 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 feature discrimination analysis supports the authors’ claims and highlights the contribution of each component in the method.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper follows a recent trend of combining CT radiology scans with vessel segmentations to guide model attention to the areas that are most relevant for aneurysm detection. Unlike previous work, which uses vessel information either as a binary mask or as a distance map, the authors extract geometric features from point clouds representing the vessels and use differentiable Gaussian splatting to generate features. Then, a dual-gated residual fusion mechanism is used to enable the model to “decide” which parts of this information to use, leading to a significant boost in detection performance. The geometric feature extraction pipeline is lightweight and depends only on the availability of 3D vessel segmentations.
- 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 contributes to bridging the gap between medical imaging and general 3D computer vision by applying 3D feature extraction and Gaussian splatting in a way that enables DL architectures to leverage vessel information in a principled way, which itself is critical for aneurysm detection. I believe this is a very clever way to integrate vessel information.
-
The proposed gating mechanism, while simple, has a positive impact and highlights the need to modulate geometric information for effective aneurysm detection.
-
The authors propose using a nnDetection model to generate candidate ROIs, simplifying the computational cost of running the model on entire scans. With more training data, this could lead to close-to-perfect filtering of relevant areas.
-
The architecture achieves a very high recall with few to no false positives, albeit this has serious caveats that I will discuss in the weaknesses section.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
There is limited clarity about how the vessel segmentation labels were obtained. The authors claim “Vessel masks are predicted solely from CTA volumes without accessing aneurysm annotations.”, which implies the authors labelled the training data not only for aneurysm detection but for vessel segmentation too. While the authors claim that they also trained the segmentation model using 5-fold cross validation, it is not clear whether the folds were a 1:1 match (ideally, each model should use vessel segmentations from a segmentation model that only had seen the exact same training data, without any access to the evaluation data for that fold). This raises concerns about data leakage as there is a non-zero chance that the areas where aneurysms occur have morphological differences present in the vessel segmentations that the downstream model could leverage to predict the existence of an aneurysm. Furthermore, there is no indication about what extent of the brain vasculature is covered in these annotations. The authors should clarify on this point.
-
The need to train a new vessel segmentation model on newly annotated data seems questionable considering there are several publicly available models dating as far back as 2023 that can segment vessels, ranging from just the Circle of Willis to almost the entire arterial network. Likewise, datasets such as TopCoW, TopBrain, and VesselVerse could have been used to train a segmentation network. In all fairness, this is not a weakness per se, but using existing data or models could have helped solve the critical concerns listed above.
-
The evaluation does not provide insight on whether the proposed method generalizes well outside of the training data, which per the authors’ own claims only contains aneurysms from a few arterial regions. The authors describe using five-fold cross-validation on a single data set with no held-out data. Essentially, this means that the models are evaluated in-domain. As of today, there exist over 400 CTA evaluation scans with dense aneurysm annotations from datasets such as the Large IA Dataset (https://zenodo.org/records/6801398) or CMHA (https://springernature.figshare.com/articles/dataset/CMHA_Intracranial_Aneurysm_CTA_Image_3D_Model_Dataset_with_Clinical_Morphological_Hemodynamic_Data/26965450), plus the RSNA dataset with over 800 aneurysms annotated as coordinates. The Large IA Dataset also contains a sizeable training partition with over 1,000 scans. Even if the authors were compute-limited, they could have used a subset of the available training and evaluation data to better evaluate their approach. I’m sharing these links in hopes you can use them to resubmit the paper if it’s rejected, but please take into account that no new results can be reported in the rebuttal.
-
No evaluation is provided to ablate the impact of the candidate ROIs. While this is a clever way to save compute, a sliding window approach could have been implemented to have a better idea of how this model would perform in a real life scenario, particularly since results are reported based on an oracle fallback for the aneurysms where the ROI detection model failed. Likewise, there is no ablation for the use of the vessel segmentation masks as-is without Gaussian splatting, meaning there is no way to tell if the contribution comes from the features or the vasculature itself.
-
There are no comparisons with respect to methods in the literature that use vessel information to enhance aneurysm detection and the only two baselines are fairly simple methods that use only voxel information.
-
- 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?
The authors make a strong methodological contribution to aneurysm detection but the study has serious flaws that make it difficult to evaluate its potential as a tool that the clinical community could benefit from. In particular, the possibility of data leakage from training the vessel segmentation model on the same data used for aneurysm detection could make the results meaningless. Even if the authors confirm that data leakage was avoided, the evaluations made are lacking, with no acknowledgement of the large amount of available annotated data that could be used to validate the model and no comparisons with respect to existing models that tackle aneurysm detection in CTA scans. I encourage the authors to explore the datasets that were shared in the review if they want to make a more compelling argument for the relevance of their work in case the paper gets rejected.
- 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 author response addresses the most critical issue and confirms that cross-validation does avoid data leakage. I’m also satisfied with the justification of training the vessel segmentation model fully in-domain.
I still feel that the authors ignore very relevant baselines proposed over the last few years. They argue that their approach is “not comparable to end-to-end vessel-aware pipelines bundled with skull stripping/TTA/component filtering and partially non-public preprocessing”. As long as these baselines can be run with the materials that are available online the comparisons remain relevant and it’s up to the authors to discuss the impact of the “non-comparable” design choices.
In any case, I find the response regarding the external validation data wholly insufficient. While it is true that some of these public datasets are not a 1:1 match for the authors’ setup, the Large Intracranial Aneurysm dataset can be accessed quite easily through a form in Zenodo, and includes dense segmentation masks for 1,000+ aneurysms across 1,000+ scans. Moreover, the CMHA dataset provides aneurysm detection masks as 3D STL surfaces that can be converted into dense masks using 3D Slicer or even Python scripts.
I’m switching my review to an accept due to the methodological contribution of the paper, but I’d strongly encourage the authors to use the additional space to add results on the Large IA dataset if the meta-reviewers deem it acceptable or at the very least to include such results in their GitHub repository (to clarify, my choice to accept is not conditioned on the authors doing this)
Review #3
- Please describe the contribution of the paper
This paper proposes a framework to address the limitations of voxel-only segmentation in CTA, where ambiguous vessel intensities make small aneurysms difficult to distinguish from normal vessels. The main contribution is the integration of point-cloud-based geometric features into a voxel backbone, allowing explicit surface geometry to be used as a prior for aneurysm segmentation.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses a clearly motivated problem: small aneurysms often resemble normal vascular bends, making intensity-only segmentation difficult. 2.The use of Gaussian splatting as a fixed-kernel point-to-voxel projection is effective, as it enables smoother and more natural embedding of geometric features than simple nearest-neighbor assignment. 3.The overall design is coherent and well motivated, with geometric feature extraction, point-to-voxel projection, and gated fusion combined in a consistent manner.
- 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.Although the use of Gaussian splatting for modality fusion is a reasonable and technically sound design choice, the underlying idea of using point-cloud geometry as a prior is already quite common, which limits the overall methodological novelty. 2.The range of comparative baselines is limited. Since the paper emphasizes geometry-aware fusion, comparisons with more geometry-aware or vascular-structure-aware methods would strengthen the claims. In particular, because SwinUNETR is included as part of the ablation study, the comparison in Table 2 is effectively reduced to nnU-Net, which makes the empirical scope appear rather narrow.
- 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 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?
This paper presents a clinically meaningful problem with a clear motivation, a well-structured method, and a reasonable ablation study. The integration of geometric and voxel-based information is technically sound, and the paper is generally well presented. However, the overall methodological novelty remains limited, as the key idea of using point-cloud geometry as a prior is not particularly new despite the reasonable use of Gaussian splatting for modality fusion. Moreover, the comparative evaluation is somewhat narrow; since SwinUNETR is included in the ablation study, the effective comparison in Table 2 is largely limited to nnU-Net. Overall, while the work is solid and clinically relevant, I find it slightly below the acceptance threshold.
- 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.
The concern regarding novelty has been addressed by clarifying the contribution of voxel re-projection using Gaussian splatting. However, the lack of sufficient comparison with existing models remains unresolved. As a result, it is still difficult to fully assess whether the proposed methodology, whose main contribution lies in integrating Gaussian splatting, is sufficiently strong.
Author Feedback
We thank R1–R3, who affirm both the methodological (“strong methodological contribution,” “clever vessel integration”) and clinical (“coherent and well motivated”) value of our work. We address each concern below.
(1) R2 — Fold alignment and absence of leakage. No data leakage. We used one fixed patient-level 5-fold split for both vessel segmentation and IA segmentation. In each fold, both models train only on the training cases and are evaluated only on the held-out cases. Vessel masks for held-out cases come from the vessel model trained only on the training cases. No held-out case enters vessel-model training, validation, hyperparameter tuning, model selection, or IA-segmentation training. Vessel labels are binary lumen annotations with no aneurysm masks. Therefore, any morphology in the predicted vessel mask is inferred from the held-out CTA itself, not memorised from training exposure. We agree §3.2 was not explicit enough and will revise accordingly.
(2) R2&R3 — What Gaussian splatting adds over direct vessel masks. A binary mask encodes only vessel location, not the surface properties (curvature, density, normals) driving discrimination (Fig. 3b: aneurysm/vessel ratios 1.23–1.27×). Our 10-D descriptors are splatted onto the voxel grid via differentiable, sub-voxel-continuous projection—smoother than nearest-neighbour voxelization, and more directly aligned with fixed geometric descriptors than generic learnable rendering [11].
(3) R2 — External validation. Public IA resources each differ in access/modality/annotation/task: Large IA is access-restricted, CMHA has CTA + 3D models but no voxel-wise IA masks, RSNA has coordinates only, CADA uses 3DRA not CTA (closed), and IntraA is hemodynamic. Our 266-lesion multi-vendor cohort, dual-expert annotation, 5-fold CV + 95% CIs supports an in-domain study. External validation is future work.
(4) R2&R3 — Vessel-aware baselines. Our method is not comparable to end-to-end vessel-aware pipelines bundled with skull stripping/TTA/component filtering and partially non-public preprocessing. We will release code, preprocessing, and hyperparameters. In Table 3, the “w/o Gating” variant is itself a vessel-aware geometric variant: being vessel-derived, the 10 descriptors inherently mix vessel-location information with surface geometry (curvature, density, normals). It already beats the Baseline (Recall 75.0→84.6, FPs halved), though this inherent mix is necessarily noisy across channels. We agree mask-as-channel is the cleanest ablation for disentangling descriptor content from vasculature presence. That said, simply re-weighting these geometric feature channels yields a substantial leap (DSC 0.591→0.770; DGRF learns heterogeneous weights—Fig. 3c: Density 13.8%, NV 13.0%, Curvature 12.1% vs ~7–9%), supporting that surface geometric content contributes beyond vessel-awareness alone. We will add this control in the camera-ready if permitted.
(5) R2 — Why train a new vessel model? With matched folds (see (1)), our in-house vessel model introduces no leakage. We prioritized internal rigour for this submission, and external vessel models are future work. Meanwhile, our framework already tolerates moderate-quality vessel masks (Dice 77.35%).
(6) R2 — Oracle fallback. All methods receive identical ROI inputs under the same 5-fold protocol. 84.6% come from the detector, with oracle covering the remaining detector misses. The evaluation is therefore fair for matched-patch segmentation comparison.
(7) R3 — Novelty. We acknowledge that using point-cloud geometry as a prior is not new in itself. Our key novelty lies in modality alignment: to our knowledge, we are the first to re-project sparse point-cloud surface features back into the dense CTA voxel grid via Gaussian splatting and jointly train this voxel-aligned geometric stream with the voxel backbone for IA segmentation.
(8) R1 — CTA + GFE design guide. CTA = Computed Tomography Angiography. Per-channel GFE formulas accompany our code.
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 fuses point cloud geometric features with voxel features via Gaussian splatting and dual-gated residual fusion for intracranial aneurysm segmentation. The approach makes sense and the non-learned geometric features are a nice touch for interpretability. R1 was positive and recommended an oral, while R2 and R3 provided more detailed technical critiques.
R2 raised what I consider the central issue: the vessel segmentation model was trained on the same dataset as the aneurysm detection model, and it is not clear whether the cross-validation folds were aligned between the two. If not, morphological cues in the vessel segmentations could leak information about aneurysm presence, which would undermine the results. This may have a reasonable explanation that the authors simply did not describe clearly enough, so I want to give them the opportunity to address it.
R2 also noted that some results rely on oracle fallback where the ROI detector failed, which makes reported performance hard to interpret at face value. R2 and R3 both noted that baselines are too narrow, with no vessel-aware aneurysm detection methods compared and Table 2 effectively reducing to nnU-Net alone.
For the rebuttal, please try to address all reviewers’ concerns, epecially the following four: 1.confirm that vessel segmentation folds exactly match detection folds with no data overlap; 2.justify the absence of vessel-based detection baselines; 3.discuss why available external datasets were not used; 4.clarify what Gaussian splatting adds over using vessel masks directly.
Point 1 is the one that matters most for my final recommendation.
- 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 is a borderline accept and I agree with all reviewers’ comments on the original submission.
The round one review turned on one question above all others: whether the vessel-segmentation model and the aneurysm model shared cross-validation folds, since a leak there would have made the reported gains meaningless. The rebuttal settles it. The authors use a single patient-level 5-fold split for both stages, train each fold’s vessel model only on that fold’s training patients, and note that the vessel labels are binary lumen masks carrying no aneurysm information, so there is no aneurysm signal to memorize. R2, who is Very confident and who had raised this exact concern at a Weak Reject, accepted the explanation and moved to Accept. R1 had already recommended acceptance independent of rebuttal. MICCAI rebuttal rules do not allow the authors to add new experiments, but R3’s concern is correct and clearly more baseline comparisons are needed although partly answered by the existing “w/o Gating” ablation. This is the major issue should be solved in the future 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.
Reject
- Please justify your recommendation.
Although the rebuttal somehow addresses the data-leakage concern and clarifies the novelty of voxel re-projection using Gaussian splatting, important concerns remain regarding insufficient comparisons with recent relevant baselines and inadequate external validation, making it difficult to fully assess the strength and impact of the proposed methodology. Therefore, the AC recommends rejection.
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.
Overall, there is consensus that the paper addresses a relevant clinical problem and proposes a technically sound geometry-aware fusion approach for aneurysm segmentation. The main concerns have been adequately clarified in the rebuttal, and I therefore recommend acceptance.
At the same time, there remain valid concerns about the limited comparison to vessel-aware baselines, the lack of external validation. I encourage the authors to address these points as much as possible in the camera-ready version. I also strongly encourage the authors to make the code and model weights publicly available for the benefit of the community.
