List of Papers Browse by Subject Areas Author List
Abstract
Intracranial aneurysms exhibit substantial variability in size, morphology, and anatomical location, limiting the availability of diverse and balanced training data for CTA-based learning tasks such as detection and segmentation. Existing synthesis and augmentation approaches either lack explicit spatial control or struggle to generate anatomically consistent aneurysms in 3D angiography volumes. We propose ReScan-IA, a spatially-adaptive 3D diffusion framework for controllable aneurysm synthesis. We frame aneurysm inpainting as a virtual re-scan, where a clinically plausible aneurysm is introduced into an existing CTA volume as if it had been captured in a repeat acquisition. ReScan-IA integrates spatially-adaptive conditioning to control aneurysm placement and morphology. A binary vessel prior promotes vascular continuity, while multiclass Circle of Willis labels are used during synthetic sampling to select anatomically corresponding vessel territories, insertion sites, and aneurysm templates. We train on 1,223 multi-center CTA volumes from the LargeIA and RSNA datasets and evaluate utility on two independent external datasets under real-only, synthetic-only, and augmented training settings. Augmenting 250 real cases with synthetic samples improves voxel-wise Dice on External A from 41.1\% to 62.2\%, while reducing false positives per case from 5.07 to 1.83 and improving recall over the real-only baseline across both external cohorts. Implementation is available at: https://github.com/9xEzreaL/ReScan-IA
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5128_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/9xEzreaL/ReScan-IA
Link to the Dataset(s)
LargeIA dataset: https://zenodo.org/records/6801398
RSNA challenge dataset: https://www.kaggle.com/competitions/rsna-intracranial-aneurysm-detection
BibTex
@InProceedings{ChuTzu_ReScanIA_MICCAI2026,
author = { Chuang, Tzu I AND Aydin, Orhun Utku AND Tanioka, Satoru AND Frey, Dietmar AND Hilbert, Adam},
title = { { ReScan-IA: A Spatially-Adaptive Diffusion Framework for Controllable 3D Intracranial Aneurysm Inpainting } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work presents ReScanIA, a spatially-adaptive 3D diffusion framework for controllable aneurysm synthesis. Where aneurysm inpainting is represented as a virtual re-scan, where a clinically plausible aneurysm is introduced into an existing CTA volume as if it had been captured in a repeat acquisition. ReScan-IA integrates spatially-adaptive conditioning to control aneurysm placement and morphology.
- 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 papers solves an important problem of data scarcity and imbalance in intracranial aneurysm analysis.
- The paper proposes a novel controllable 3D diffusion framework with spatially-adaptive conditioning.
- The method incorporates anatomical information (vessel masks) ensuring vascular continuity.
- Experimental results shows strong improvement in segmentation and detection using synthetic data.
- Strong validation on large multi-center data and external datasets, showing good method generalization.
- 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.
- It lacks comparison with recent diffusion/GAN-based medical synthesis methods. -Limitations and Clinical Integration are not discussed adequately.
- Large descriptive blocks make the methodology hard to follow. Add clearer pipeline diagrams. -Statistical comparisons (t-test ) would enhance the proposed work validation. -Please address generalizability of the proposed to different datasets and protocols in the limitations.
- Patch-based approach may miss global anatomical consistency and shows inconsistent results across datasets.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- It lacks comparison with recent diffusion/GAN-based medical synthesis methods. -Limitations and Clinical Integration are not discussed adequately.
- Large descriptive blocks make the methodology hard to follow. Add clearer pipeline diagrams. -Statistical comparisons (t-test ) would enhance the proposed work validation. -Please address generalizability of the proposed to different datasets and protocols in the limitations.
- Patch-based approach may miss global anatomical consistency and shows inconsistent results across datasets.
- 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 presents ReScan-IA, a 3D diffusion-based framework for controllable aneurysm synthesis in CTA, addressing the limited availability and diversity of annotated data. The method formulates aneurysm synthesis as a virtual re-scan process, using masked diffusion inpainting to generate aneurysms within existing volumes. A key contribution is the use of spatially-adaptive conditioning together with vessel priors to control aneurysm location and morphology while ensuring anatomical consistency. Conditioning on Circle of Willis segmentations and vessel masks constrains synthesis to valid vascular regions and preserves vessel continuity. The method is evaluated for data augmentation across multiple training settings and external datasets, and demonstrates improved performance in downstream segmentation and detection tasks. This approach provides a practical and clinically relevant solution for controllable medical image synthesis in aneurysm analysis.
- 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 problem, namely the limited availability of annotated aneurysm data in CTA, which restricts the performance of deep learning models. A key strength is the integration of controllable generative modeling with anatomical priors. Spatial control signals and vessel information guide synthesis and help ensure anatomically consistent results. The use of the method for data augmentation is well motivated. Improvements in downstream segmentation and detection tasks support its practical value. The use of 3D volumetric modeling is appropriate and avoids the limitations of 2D approaches.
- 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 paper has several weaknesses that reduce clarity and limit the strength of the experimental evaluation. 1.The model architecture is under-specified in the manuscript. Key details such as the number of U-Net levels, channel sizes, normalization layers, and the exact implementation of the 3D SPADE-like modulation are not clearly described. Although code is provided, these details should also be documented in the paper. 2.The training and validation protocol is not clearly described. The paper does not specify the train/validation/test split ratio, whether splitting is performed at the patient level, or how the LargeIA and RSNA Intracranial Aneurysm Detection challenge datasets are combined. This makes it difficult to assess experimental rigor and potential data leakage. 3.The external test sets are insufficiently described. While two independent cohorts (External A and External B) are used, the number of samples in each dataset is not reported. In addition, their source and acquisition characteristics are not described. Without these details, the strength of the cross-dataset generalization claims is difficult to assess. 4.The patch-based training strategy is unclear. The paper does not explain how patches are sampled (random vs. aneurysm-centered), how overlap is handled during inference, or how full volumes are reconstructed. Potential boundary artifacts are not discussed. 5.The ablation study is limited and lacks quantitative comparison. The analysis relies mainly on qualitative results. There is no systematic quantitative evaluation of key components such as vessel prior (v), spatial control (s), or their interaction. 6.Preprocessing choices are not justified. Key parameters such as resolution, intensity clipping are described but not motivated or analyzed for sensitivity. Post-processing steps are also not described.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 paper addresses an important problem and presents a promising approach. To further strengthen the work, the authors should improve clarity and completeness of the experimental description, including dataset reporting and evaluation setup. Providing more implementation details in the paper would improve reproducibility, even if code is available. Expanding the ablation study with quantitative results would also help better support the contribution of each component. Finally, including analysis of failure cases and a clearer discussion of controllability and clinical applicability would improve the overall impact.
- 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 addresses a clinically relevant problem, namely the limited availability of annotated aneurysm data in CTA, and proposes a practical approach using controllable 3D diffusion models. A key strength is the use of spatial control and anatomical priors to guide synthesis, which helps ensure anatomically consistent results. The formulation is well motivated, and the use of 3D modeling is appropriate. The reported improvements in downstream segmentation and detection tasks, including evaluation on external datasets, support the practical value of the method. There are some limitations in the clarity of the experimental setup and reporting. In particular, dataset composition, data splits, and details of the external test sets are not fully described. The model architecture and training strategy could be documented more clearly in the manuscript, and the ablation study would benefit from additional quantitative analysis. Some preprocessing choices are also not fully justified. These issues mainly affect clarity and completeness rather than the core contribution.
- 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 #3
- Please describe the contribution of the paper
The authors propose using both a vessel mask and the aneurysm mask to synthesize realistic aneurysms in CTA images. They adapt SPADE layers to 3D for the mask-based conditioning. The idea of using a vessel mask in addition to the aneurysm location enables generating CTA with vascular connectivity and adherence to the morphology.
- 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.
Motivation: The idea of using spatial masks to ensure vascular structure, rather than focusing only on the aneurysm mask, is very neat.
Adapting SPADE layers: SPADE layers are designed to condition on segmentation masks. Extending them to 3D and showing superior results (see minor points) suggests an alternative to pure concatenation and expensive cross-attention-based conditioning.
- 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.
Ambiguity: A few points in the paper are unclear. For instance, the authors mention that they use the multi-class Circle of Willis labels to get the vascular structure. However, it appears that the multi-class labels are not used; it is simply a binary segmentation mask.
Experimental Setup: The experiments are not clear to me. For instance, in the segmentation-only case (2), 25 locations were sampled for each of the 10 patients. The locations should be sampled only from the existing patients (i.e, the 10 patients), else it is an unfair comparison. I am not sure if that was the case.
Further, the 250 synthetic samples seem to work better than the 250 real samples. So, the synthetic samples are “more diverse” than the real samples. This makes the last experiment completely uninutitive to me. If you can get so much out of just 10 real samples, then the same experiment should be repeated for the 250 samples, and you would get even better results. Why stick with the 250 synthetics generated from 10 samples?
How are the cases with multiple aneurysms handled? The RSNA Intracranial Aneurysm Detection challenge shows that patients can have multiple aneurysms. In such a scenario, do the authors always select all the aneurysms or only a subset?
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 experiments without Spatially Adaptive Modulation (i.e., SPADE layers), do the authors use a simple concatenation-based conditioning, as for the vascular masks, or some other conditioning strategy?
Architecture: In equation 3, the noisy masked region is also appended to the background channel. Is there some ablation to show that including this background channel explicitly makes the learning easier?
The authors should also provide the distribution of the number of aneurysms and their locations to give an idea of the diversity in the training dataset.
- 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?
I like the general idea of using additional vascular segmentation conditioning to ensure that the generated samples are realistic. The idea of extending SPADE layers to 3D is also very interesting. However, the experiment setup is not very clear to me. I would like the authors to address some of the concerns I raised here about their experiments.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers for recognizing the strengths of our work and for the early acceptance decision. We would like to address a few points raised by the reviewers briefly.
Experimental setup, patient-level split, and leakage prevention. We will revise the manuscript to clarify the patient-level split and the relationship between ReScan-IA training, synthetic generation, downstream training, and external evaluation datasets. The ReScan-IA training set consisted of 1223 subjects, including 287 from the RSNA dataset and 936 from the LargeIA dataset. An additional 250 held-out LargeIA subjects formed the internal pool for downstream aneurysm segmentation experiments. External datasets A and B were not used for either ReScan-IA diffusion training or downstream model training. For downstream segmentation and detection, all real subjects came from the held-out internal downstream experiments training dataset. Synthetic aneurysm templates were sampled only from the ReScan-IA training pool, excluding External A/B. This increased lesion diversity without patient-level leakage.
Circle of Willis labels and binary vessel prior. We will revise the wording to clearly distinguish the binary vessel prior from multi-class vessel location control. In our implementation, the diffusion model receives only a binary vessel prior as input-level anatomical conditioning. Multi-class Circle of Willis labels are used only during synthetic sampling to identify the parent vessel, select anatomically corresponding insertion sites, and match aneurysm templates by vessel territory. This was done to improve the anatomical plausibility of introduced synthetic aneurysms.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
despite of a weak reject with reasonable justifications, the reviews are still good enough to support an early accept.