List of Papers Browse by Subject Areas Author List
Abstract
Visceral artery aneurysms (VAAs) are relatively rare but potentially life-threatening vascular dilatations that are frequently detected incidentally on abdominal CT, making opportunistic screening on non-contrast CT (NCT) clinically important. However, automated VAA detection on NCT is fundamentally challenging due to weak vessel–tissue contrast, complex vascular branching, and the continuous transition between normal vessel wall and aneurysmal dilation. We reformulate VAA screening as a dual-prior representation learning problem rather than a pure segmentation task. Built upon a multi-label vascular segmentation backbone, the proposed DeepVAA jointly learns two complementary priors defined in continuous target spaces: a geometric confidence prior that models aneurysmal dilation as a smooth structural response along the vessel wall, and an anatomical reliability prior that encodes geometry-induced ambiguity, such as bifurcations and high-curvature segments, as a continuous penalty field. During inference, the learned reliability prior explicitly calibrates the confidence map through structured prior projection and multiplicative suppression, yielding a prior-calibrated lesion probability. Extensive experiments on multi-center cohorts demonstrate consistent performance improvements and strong cross-site generalization, highlighting the robustness and effectiveness of the proposed dual-prior framework for opportunistic VAA screening on NCT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2314_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaJia_Learning_MICCAI2026,
author = { Zhang, Jianfeng AND Liu, Tianyi AND Bai, Long AND Xu, Zhuowei AND Hu, Yujian AND Zhou, Yide AND He, Yangyan AND Wang, Zirui AND Zhou, Min AND Zhang, Hongkun AND Xu, Minfeng},
title = { { Learning Dual Prior Presentation for Opportunistic Screening of Visceral Artery Aneurysms on Non-Contrast CT } },
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
This paper addresses opportunistic screening of visceral artery aneurysms (VAAs) on non-contrast CT, a clinically important but technically difficult problem due to weak vessel–tissue contrast, small lesion size, and confounding vascular geometry. The authors propose DeepVAA, a dual-prior representation learning framework built on top of a multi-label vascular segmentation backbone. Instead of treating VAA detection purely as a discrete segmentation problem, the method learns two continuous target spaces: a geometric confidence prior to model aneurysmal dilation and an anatomical reliability prior to suppress false activations in ambiguous regions such as bifurcations and tortuous vessel segments. During inference, the reliability prior calibrates the confidence prior through multiplicative suppression. The paper reports consistent gains over several strong 3D segmentation baselines on both internal and external cohorts.
- 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.
A major strength is the clinical relevance. Opportunistic screening from routine CT is an important direction in medical imaging, and prior work has highlighted opportunistic CT screening as a meaningful translational use case. The paper also targets a problem where prior AI work appears limited, with at least one notable earlier VAA detection study focusing on CTA, not non-contrast CT, which makes the present setting more challenging and potentially impactful. The key idea of replacing a pure discrete lesion class formulation with two continuous prior spaces is well motivated by the anatomy and radiology of VAAs. The geometric confidence prior matches the notion of gradual dilation, while the reliability prior explicitly models anatomically ambiguous regions that commonly trigger false positives. This is more thoughtful than simply adding another segmentation head, and the page-3 framework diagram makes the design easy to understand. In Table 1, DeepVAA improves over nnUNet, U-Mamba, SegMamba, MedNeXt, SwinUNETR, and UXNet on the reported patient-level F1 and VAA Dice metrics, and the external cohort results suggest that the method retains performance better than several baselines under site shift. The ablation tables on page 7 also support that the two priors contribute differently and complement one another: the geometric branch mainly boosts recall, while the anatomical reliability branch mainly improves precision.
- 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.
My biggest concern is the insufficient characterization of the dataset and task difficulty. The paper reports 1000 internal and 100 external cases, but it does not clearly state how many patients actually had VAA, how many aneurysms were present, the lesion size distribution, the breakdown by arterial territory, or the prevalence in the internal and external cohorts. Without those details, it is hard to assess the true screening difficulty and the clinical meaning of the reported metrics. The paper reports the proposed method at multiple thresholds (0.05 to 0.30), while the baselines are shown at a single operating point. This makes the comparison somewhat hard to interpret unless the threshold was fixed a priori on validation data and then frozen for all testing. A stronger presentation would report PR curves, FROC-style analysis, or a clearly pre-specified threshold selection procedure. In addition, the lesion-level true-positive definition only requires at least one voxel of overlap, which is permissive and may overestimate lesion detection quality for tiny lesions. Most comparisons are against generic 3D medical segmentation architectures rather than methods tailored to aneurysm detection or vascular screening. The paper cites prior AI-based VAA detection on CTA, but does not compare conceptually or experimentally to simpler diameter-based, centerline-based, or rule-based vascular abnormality detectors. Since earlier work has already explored automated VAA detection on CTA, a clearer comparison to task-specific alternatives would strengthen the novelty and clinical argument. I am not fully satisfied with the clinical validation framing. The paper uses registered CTA-derived annotations projected onto NCCT, which is practical, but the manuscript would be stronger if it clarified registration quality control, inter-reader agreement, and whether the external cohort was collected with materially different scanners or protocols. In a clinically oriented screening paper, these details matter for judging translational robustness.
- 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?
The study uses retrospective human imaging data from two hospitals and manual expert annotations, but I did not see an explicit IRB / ethics approval or consent-waiver statement in the submitted manuscript. This may simply be omitted for anonymity or space, but I would want it clarified.
- 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?
I lean slightly positive because the paper tackles a clinically meaningful problem that appears underexplored in the literature, especially for non-contrast CT, and the proposed dual-prior formulation is more than a trivial architectural tweak. The empirical results are consistently good, the qualitative examples are supportive, and the external validation cohort is a real strength. Prior work has demonstrated automated VAA detection on CTA, but this paper targets a harder, more opportunistic NCT setting, which adds practical value. That said, I am only at weak accept because the paper still has several important gaps: incomplete cohort characterization, potentially optimistic evaluation choices around thresholding and lesion matching, limited task-specific baselines, and moderate reproducibility. If the rebuttal clearly addresses these concerns, especially the evaluation protocol and data description, I would be comfortable keeping the paper above threshold.
- 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 proposes a dual‑prior representation learning framework designed to enhance the segmentation of vascular structures in non‑contrast CT scans. Built on top of a multi‑label vascular segmentation backbone, the method introduces two complementary priors that separately capture dilation‑sensitive responses and geometry‑driven activations. By integrating these priors into a unified segmentation architecture, the model aims to improve the detection and screening of vascular abnormalities.
The approach is evaluated on two datasets from different hospitals, including a large internal training cohort (1,000 cases) and an external test set (100 cases). The results demonstrate that the proposed framework achieves stronger performance than several established segmentation models, highlighting its potential for improving non‑contrast CT–based vascular screening.
- 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 separation of dilation‑sensitive features from geometric structure representations introduces a meaningful and well‑motivated architectural innovation. This helps the model emphasise both morphological integrity and pathological dilation patterns, which are clinically relevant in vascular disease.
Non‑contrast CT is widely available and commonly used in acute care. Demonstrating improved vascular segmentation without contrast agents enhances potential applicability in real‑world clinical environments.
Training on one hospital dataset and testing on another provides evidence of cross‑site generalisation, which is an important step toward clinical robustness.
- 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 framework is technically sound, the paper does not address how the dual priors influence the final segmentation or how clinicians can interpret the resulting predictions. Tools such as attention visualisation, relevance maps, or uncertainty estimation would help demonstrate whether the model focuses on clinically meaningful vascular regions. Without these, the transparency of the method remains limited.
Although two hospital datasets are used, the external testing cohort is relatively small (100 cases), and the study does not evaluate performance across different CT acquisition parameters, noise levels, and scanner variations. Non‑contrast CT imaging is subject to substantial variability in practice, and without robustness testing it is unclear how well the model generalises to broader populations.
- 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
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 proposes a well‑motivated dual‑prior segmentation framework and demonstrates solid improvements over standard baselines. The methodological ideas are interesting, and the application to non‑contrast CT is clinically meaningful.
- Reviewer confidence
Somewhat confident (2)
- [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.
N/A
Review #3
- Please describe the contribution of the paper
– The authors identify three core technical challenges in VAA detection on NCT, including low vessel-tissue contrast, geometric ambiguity at vascular bifurcations and high-curvature segments, and data scarcity due to the rarity of the disease, and propose a dual-prior representation learning framework (DeepVAA) designed to address these challenges. – Two complementary priors are introduced: a geometric confidence prior (GCP) that models aneurysmal dilation as a continuous structural response along the vessel wall, and an anatomical reliability prior (ARP) that encodes geometry-induced ambiguity as a continuous penalty field. During inference, the ARP calibrates the GCP-derived confidence map through structured prior projection and multiplicative suppression.
- 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 reformulation of VAA detection as a representation learning problem offers a conceptually distinct formulation from conventional binary segmentation approaches, and the dual-prior design is motivated by the specific challenges of the task. – The framework attempts to simultaneously address lesion detection sensitivity and false positive suppression through two complementary priors, which is a reasonable design consideration given the anatomical complexity of the problem. – The inclusion of an external cohort for cross-site validation reflects an intention to assess generalizability beyond the training domain, which is a step toward broader validation.
- 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.
– Both GCP and ARP are generated through hand-crafted rules rather than being learned directly from data. This raises a fundamental question about whether the framework truly constitutes “representation learning” as claimed, or whether it is better characterized as rule-based feature injection. Furthermore, the combination of the two priors at inference via simple multiplicative suppression is a relatively rudimentary operation that limits the expressive capacity of the overall framework. – Although the model takes NCT as input, the ground truth labels are derived from CTA. Given that NCT and CTA differ fundamentally in image characteristics, vascular boundaries that are clearly delineated on CTA may not have corresponding visual information on NCT. The authors do not provide any analysis of how registration errors between the two modalities may have affected training, nor do they discuss whether the model may have been trained to detect boundaries that are physically invisible on NCT. This inconsistency appears to be in direct conflict with the premise of opportunistic screening on NCT. – Table 1 demonstrates considerable sensitivity of performance metrics to the choice of threshold tau, yet no objective criterion or automated mechanism for selecting the optimal value is provided. The current approach of selecting tau based on internal cohort performance raises concerns about post-hoc tuning, and it remains unclear whether the chosen threshold would generalize to external sites or different imaging protocols. – The external cohort consists of only 100 cases, and neither the internal nor external cohort reports the number of VAA-positive patients or individual lesions. Given the rarity of VAA, the actual number of positive samples is likely very small, which undermines the statistical reliability of the reported performance metrics. Additionally, the absence of size-stratified analysis makes it impossible to assess whether the model performs robustly across the full spectrum of aneurysm sizes. – While ARP is designed to suppress false positives, false negatives carry greater clinical risk in a screening setting. The observed decline in recall from 78.2% to 77.4% upon addition of ARP, though numerically modest, suggests that some true lesions may have been suppressed in the process. Critically, there is no error analysis characterizing which lesions were missed, whether small, located near bifurcations, or exhibiting atypical morphology, nor any quantitative comparison of prediction maps before and after ARP application. – The paper positions DeepVAA as a tool for opportunistic screening, yet provides no discussion of how it would integrate into actual clinical workflows. With precision ranging from 80% to 86%, approximately 10–20% of model-positive cases would be false positives, each potentially triggering unnecessary follow-up CTA. For a screening tool targeting an asymptomatic population, this false positive burden requires explicit clinical justification. The manuscript does not address what level of false positive rate would be acceptable in practice, nor does it discuss the cost-benefit implications of deploying such a system in routine radiology workflow.
- 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?
The submission appears to involve human/clinical data, but I could not identify any statement regarding IRB/ethics committee approval, informed consent, or waiver of consent. This raises a potential ethics concern that should be clarified by the authors.
- 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 addresses a challenging and clinically relevant problem and the proposed dual-prior representation learning framework offers a conceptually distinct formulation from conventional segmentation approaches. However, as detailed in the weaknesses above, several fundamental concerns make it difficult to support acceptance in its current form. These include a methodological inconsistency between the NCT-based inference premise and CTA-derived ground truth generation, a gap between the claimed representation learning framework and its largely heuristic implementation, limited external validation with insufficient reporting of positive sample statistics, and the absence of false negative analysis and clinical workflow consideration, both of which are critical for evaluating the practical utility of a screening tool. These issues warrant careful revision and additional experiments. The reviewer looks forward to the authors’ response in the rebuttal.
- 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
Author Feedback
We thank the reviewers for their valuable feedback and for recognizing our work as clinically meaningful (R1) and well motivated (R2). We address the major concerns below:
Dataset Setup (R1 R2 R3) Each set (train, valid & external) follows a negative/positive patient ratio of 1:1±0.02. The three sets contain 563, 124 & 81 lesions, respectively (celiac trunk 79%, superior mesenteric 11%, right renal 6%, left renal 4%). The recall increased with lesion size, ranging from 61% (<10 mm), 83% (10-20mm), to 100% (>20 mm); these size groups accounted for 33.8%, 61.8% & 4.4%. The internal cohort has two scanner brands (GE, UIH), and the external cohort has four (+Philips, Siemens). Our high performance strongly demonstrates our cross-center robustness. We acknowledge that the external set is relatively small, but our dataset is already the largest NCT-CTA cohort with fine-grained VAA annotations. We will also further expand our multi-center datasets.
Tau selection & FP processing (R1 R3) DeepVAA achieves patient AUROC 0.9306 and lesion PRAUC 0.9266.Threshold 0.3 was predefined on validation data, maximizing the Youden Index (0.8365) at the ROC elbow (Sensitivity 86.5%, FPR 2.9%). Clinically, DeepVAA is a triage tool. Size-based analysis shows 14/16 FPs are <15mm artifacts easily dismissed by radiologists, avoiding unnecessary CTAs. Crucially, high-risk VAAs (>20mm) achieve 100% recall with no FPs.
Dual Representation Prior (R2 R3) We have addressed this in the ablation study. In Sec 3.2, the dual priors are effective not only when applied individually, but also when working jointly. We respectfully clarify that GCP and ARP are explicitly implemented as learnable convolutional branches optimized by auxiliary losses. Consequently, the network genuinely learns these anatomical representations, and their multiplication at inference acts as a dynamic, data-driven spatial attention mechanism rather than a primitive rule injection.
Registration Quality (R1 R3) Offline CTA was used only to overcome the difficulty of manual annotation on NCT. This cross-modality label transfer is a mature paradigm [PANDA & iAorta; Nat Med 2023, 2025]. Because deployment relies exclusively on NCT, the approach is fully compatible with opportunistic screening. Besides, NCT & CTA were acquired consecutively under strict immobilization, and visceral arteries are largely motion-stable. Visual QC confirmed that DeepVAA does not hallucinate CTA boundaries, but learns objective NCT signs including perivascular fat displacement, contour bulging, and calcifications.
Recall Drop with ARP (R3) In the validation set (124 lesions), the slight recall drop was due to one missed 7 mm VAA. Given the substantial reduction in FP, missing a <10 mm VAA could be acceptable, as <10 mm lesions are generally not clinically critical cases.
Embed with Clinical Workflow (R2 R3) DeepVAA is designed as a PACS-integrated screening assistant rather than a standalone diagnostic system. It runs in the background on routine CTs, providing patient flags and lesion-level visual overlays as a secondary reader without interrupting standard workflows. For AI-flagged cases, the radiologist remains the final decision-maker, verifying highlighted candidates and recommending CTA only when clinically warranted, enabling practical workflow integration.
Compare with Vascular Abnormal Detection (R3) We also tested the centerline-based method that measured cross-sectional diameter and detected VAAs via focal diameter enlargement. It suffered a severe trade-off (70% recall/46% specificity vs 33% recall/78% specificity). In contrast, DeepVAA achieves 86.5% recall and 97.1% specificity. Such rule-based methods depend on rigid geometric priors and are limited by the complex branching anatomy. In contrast, DeepVAA learns continuous 3D anatomical context rather than relying on brittle geometric rules.
All suggested changes will be added to the paper. The code and IRB info will be public upon acceptance.
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 addresses a clinically important and under-explored task: opportunistic visceral artery aneurysm (VAA) screening on non-contrast CT. Two reviewers support weak accept; one recommends weak reject due to evaluation and clinical validity concerns. All agree that the core idea is sound and clinically valuable. 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.
Authors addressed the main concerns of the Reviewers. Although still the reproducibility issue stands, authors state in their rebuttal having obtained approval for release of the data and code that allow reproducibility of the results, which will be provided upon acceptance.
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.
based on the contributions from both clinical and technical sides, I am inclined to accept it.
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.
The reviewers generally agree that this paper addresses a clinically important and under-explored problem. The rebuttal provides useful clarification on most concerns. I therefore recommend acceptance.