Anatomy-Informed Deep Learning for Abdominal Aortic Aneurysm Segmentation

Babette Dellen; Martin Br\"uckmann; Osamah Sufyan; Ralph Wickenh\"ofer; Uwe Jaekel

arxiv: 2604.10312 · v1 · submitted 2026-04-11 · 💻 cs.CV · cs.LG

Anatomy-Informed Deep Learning for Abdominal Aortic Aneurysm Segmentation

Osamah Sufyan , Martin Br\"uckmann , Ralph Wickenh\"ofer , Babette Dellen , Uwe Jaekel This is my paper

Pith reviewed 2026-05-10 15:31 UTC · model grok-4.3

classification 💻 cs.CV cs.LG

keywords abdominal aortic aneurysmsegmentationU-Netanatomical priorsorgan exclusion masksCT angiographydeep learningfalse positives

0 comments

The pith

Organ exclusion masks improve U-Net accuracy for abdominal aortic aneurysm segmentation on limited data

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to improve segmentation of abdominal aortic aneurysms in CT angiography by adding anatomical knowledge to a standard deep learning model. Nearby organs with similar image intensities often cause false positives, and small training datasets make the problem worse. The authors create organ exclusion masks that mark non-vascular tissue and penalize the model for predicting aneurysms inside those masks during training. This guides the U-Net to focus on the aorta and its dilation while suppressing implausible outputs. The result is higher accuracy, fewer false positives, and smoother boundaries than a plain U-Net even when data is scarce.

Core claim

The anatomy-aware segmentation framework integrates organ exclusion masks derived from TotalSegmentator into U-Net training. These masks identify non-vascular organs and penalize aneurysm predictions within those regions, guiding the model to suppress anatomically implausible outputs. As a result, the model attains high accuracy, substantially reduces false positives, and improves boundary consistency compared to a baseline U-Net, even when trained on a relatively small dataset.

What carries the argument

Organ exclusion masks that identify non-vascular organs and penalize predictions inside those regions during U-Net training

If this is right

High segmentation accuracy is reached despite using a relatively small training dataset.
False positives from nearby organs are substantially reduced.
Boundary consistency of the segmented aneurysm improves over a standard U-Net.
Anatomical priors through exclusion masks enhance robustness and generalization for this task.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same mask-based prior approach could help segment other vessels or structures near tissues of similar appearance.
Explicit anatomical constraints may allow deep learning models to succeed with smaller medical imaging datasets in general.
Clinical tools for aneurysm monitoring might become more reliable if this method is integrated into existing pipelines.
The technique could be tested on MRI or other modalities to check whether the benefit holds beyond CT angiography.

Load-bearing premise

The organ exclusion masks accurately identify non-vascular organs without incorrectly excluding or overlapping regions that contain the aneurysm or its boundaries.

What would settle it

Test scans where the aneurysm borders an excluded organ and the anatomy-aware model shows no drop in false positives or begins missing aneurysm parts would challenge the value of the masks.

Figures

Figures reproduced from arXiv: 2604.10312 by Babette Dellen, Martin Br\"uckmann, Osamah Sufyan, Ralph Wickenh\"ofer, Uwe Jaekel.

**Figure 2.** Figure 2: Pipeline of the proposed method from preprocessing to training. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: (a) Qualitative comparison showing two representative slices from different [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: 3D reconstructed aneurysm from predicted segmentation (right) and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

read the original abstract

In CT angiography, the accurate segmentation of abdominal aortic aneurysms (AAAs) is difficult due to large anatomical variability, low-contrast vessel boundaries, and the close proximity of organs whose intensities resemble vascular structures, often leading to false positives. To address these challenges, we propose an anatomy-aware segmentation framework that integrates organ exclusion masks derived from TotalSegmentator into the training process. These masks encode explicit anatomical priors by identifying non-vascular organsand penalizing aneurysm predictions within these regions, thereby guiding the U-Net to focus on the aorta and its pathological dilation while suppressing anatomically implausible predictions. Despite being trained on a relatively small dataset, the anatomy-aware model achieves high accuracy, substantially reduces false positives, and improves boundary consistency compared to a standard U-Net baseline. The results demonstrate that incorporating anatomical knowledge through exclusion masks provides an efficient mechanism to enhance robustness and generalization, enabling reliable AAA segmentation even with limited training data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper applies TotalSegmentator-derived organ masks as exclusion priors during U-Net training for AAA segmentation in CT angiography, but the abstract gives no metrics or validation details to back the reported gains.

read the letter

The main takeaway is that the authors take a pre-trained TotalSegmentator model, generate masks for nearby organs, and add a penalty term to the loss so the U-Net avoids predicting aneurysm inside those regions. This is positioned as a way to cut false positives and improve boundaries when training data is limited. The specific pairing of TotalSegmentator masks with AAA segmentation is new enough in the literature on this narrow task, though it builds directly on existing anatomy-informed segmentation work rather than introducing a new principle. The motivation is clear and practical: CT angiography of AAAs often has intensity overlap with adjacent structures, and explicit priors can help a small-data model stay anatomically plausible. That part of the idea is straightforward and worth considering for similar vascular problems. The abstract claims high accuracy, fewer false positives, and better boundary consistency than a plain U-Net, which aligns with what the method should do if the masks are reliable. However, no numbers appear—no Dice scores, no dataset size, no cross-validation details, no error bars. That absence makes it impossible to judge whether the gains are real or modest. The stress-test concern also lands: TotalSegmentator was trained on non-pathological scans, yet large AAAs displace organs like the duodenum or IVC, so the exclusion masks could overlap true aneurysm voxels or boundaries. Without reported overlap statistics or an ablation that erodes the masks, the penalty term risks suppressing valid predictions and undercutting the claimed improvements. This work is aimed at medical-image segmentation researchers who already use anatomical priors or work on vascular CT tasks. A reader looking for a concrete, implementable trick to stabilize small-data training would get the method description and could try it, but would still need the results tables to decide on adoption. The paper deserves peer review. The core mechanism is described cleanly and the clinical problem is relevant; referees can check the experiments, request the missing overlap analysis, and assess whether the quantitative evidence holds up.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes an anatomy-aware segmentation framework for abdominal aortic aneurysms (AAAs) in CT angiography. It augments a U-Net with organ exclusion masks from TotalSegmentator, which are used to add a penalty term during training that discourages predictions inside non-vascular organs. The central claim is that this yields high accuracy, substantially fewer false positives, and better boundary consistency than a standard U-Net, even when trained on a relatively small dataset.

Significance. If the empirical results and mask-validity assumptions hold, the work would illustrate a lightweight, reproducible way to inject explicit anatomical priors into medical segmentation networks. This could be valuable for improving specificity and generalization in data-limited settings where intensity-based methods struggle with adjacent structures.

major comments (3)

[Abstract] Abstract: The claims of 'high accuracy', 'substantially reduces false positives', and 'improves boundary consistency' are presented without any quantitative metrics (Dice, sensitivity, specificity, Hausdorff distance), dataset size, train/test split, validation protocol, error bars, or statistical tests. This leaves the central performance claim without visible supporting evidence.
[Methods] Methods section describing the loss and mask integration: The penalty term improves specificity without harming sensitivity only if the TotalSegmentator exclusion masks are strictly disjoint from true AAA voxels and boundaries. No overlap statistics (e.g., IoU between masks and manual AAA ground truth), sensitivity analysis (erosion/dilation of masks), or validation on aneurysmal cases are reported, despite TotalSegmentator being trained on non-aneurysmal CTs where large AAAs can displace organs.
[Results] Results section: No baseline comparison details, exact metric values, number of test cases, or cross-validation scheme are provided to substantiate superiority over the standard U-Net, making it impossible to evaluate the reported gains in accuracy and boundary consistency.

minor comments (2)

[Abstract] Abstract contains a typographical error: 'non-vascular organsand' should read 'non-vascular organs and'.
The manuscript would benefit from a figure showing example TotalSegmentator masks overlaid on AAA cases to visually demonstrate the spatial relationship at vessel-organ interfaces.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thorough and constructive review. The comments highlight important areas for improving clarity and rigor, particularly regarding quantitative support for claims and validation of the anatomical priors. We address each point below and will revise the manuscript accordingly to strengthen the presentation of our anatomy-informed framework.

read point-by-point responses

Referee: [Abstract] Abstract: The claims of 'high accuracy', 'substantially reduces false positives', and 'improves boundary consistency' are presented without any quantitative metrics (Dice, sensitivity, specificity, Hausdorff distance), dataset size, train/test split, validation protocol, error bars, or statistical tests. This leaves the central performance claim without visible supporting evidence.

Authors: We agree that the abstract should provide quantitative support for the performance claims. In the revised manuscript, we will expand the abstract to include key metrics (e.g., Dice coefficient, sensitivity, specificity, Hausdorff distance) with error bars, dataset size, train/test split details, and mention of statistical testing. These results are already computed and presented in the Results section; the abstract will now reference them directly to make the central claims evidence-based. revision: yes
Referee: [Methods] Methods section describing the loss and mask integration: The penalty term improves specificity without harming sensitivity only if the TotalSegmentator exclusion masks are strictly disjoint from true AAA voxels and boundaries. No overlap statistics (e.g., IoU between masks and manual AAA ground truth), sensitivity analysis (erosion/dilation of masks), or validation on aneurysmal cases are reported, despite TotalSegmentator being trained on non-aneurysmal CTs where large AAAs can displace organs.

Authors: This point is well taken and identifies a genuine gap in the current Methods description. We will add overlap statistics (IoU between TotalSegmentator masks and manual AAA ground truth) computed on our dataset. We will also include a sensitivity analysis varying mask boundaries via erosion and dilation. For aneurysmal cases, we will explicitly discuss the limitation that TotalSegmentator was trained on non-aneurysmal data and may be affected by organ displacement; we will add qualitative assessment on our aneurysmal test cases and note this as a limitation. revision: partial
Referee: [Results] Results section: No baseline comparison details, exact metric values, number of test cases, or cross-validation scheme are provided to substantiate superiority over the standard U-Net, making it impossible to evaluate the reported gains in accuracy and boundary consistency.

Authors: We acknowledge the need for greater transparency in the Results section. The revised manuscript will include exact metric values (Dice, sensitivity, specificity, Hausdorff distance) with standard deviations and statistical significance tests comparing the anatomy-informed model to the standard U-Net baseline. We will specify the number of test cases, the precise train/test split, and the cross-validation protocol. Expanded tables and additional qualitative figures will illustrate the reduction in false positives and boundary improvements. revision: yes

Circularity Check

0 steps flagged

No circularity; external priors and baseline comparison are independent

full rationale

The derivation relies on an external pre-trained tool (TotalSegmentator) to produce organ exclusion masks, which are then used to add a penalty term during U-Net training. This is directly compared against a standard U-Net baseline on the same dataset. No equations or steps reduce by construction to fitted inputs, no self-citation chain supports the central claim, and no uniqueness theorem or ansatz is imported from prior author work. The method is self-contained against external benchmarks and does not rename known results or smuggle assumptions via self-reference.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the accuracy of an external pre-trained segmentation tool and the assumption that excluding its predicted regions will not remove valid aneurysm tissue.

axioms (1)

domain assumption Organ exclusion masks from TotalSegmentator accurately delineate non-vascular organs without overlapping the abdominal aorta or aneurysm.
Invoked when the masks are used to penalize predictions; if false, the guidance could remove true positive regions or leave false positives unaddressed.

pith-pipeline@v0.9.0 · 5472 in / 1242 out tokens · 61877 ms · 2026-05-10T15:31:21.822112+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

[1]

Systematic Review and Meta-Analysis of the Incidence of Rupture, Repair, and Death of Small and Large Abdominal Aortic Aneurysms under Surveillance

Nicola Leone et al. “Systematic Review and Meta-Analysis of the Incidence of Rupture, Repair, and Death of Small and Large Abdominal Aortic Aneurysms under Surveillance”. In:Journal of Clinical Medicine12.21 (2023), p. 6837.doi:10.3390/jcm12216837

work page doi:10.3390/jcm12216837 2023
[2]

Yamagishi, N

Taehun Kim et al. “Computed Tomography-Based Automated Measurement of Abdominal Aortic Aneurysm Using Semantic Segmentation with Active Learning”. In:Scientific Reports14(2024), p. 8924.doi:10.1038/s41598- 024-59735-8

work page doi:10.1038/s41598- 2024
[3]

U-Net: Convolu- tional Networks for Biomedical Image Segmentation

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. “U-Net: Convolu- tional Networks for Biomedical Image Segmentation”. In:Medical Image Computing and Computer-Assisted Intervention (MICCAI). 2015, pp. 234– 241

work page 2015
[4]

A Deep Learning System for the Automated De- tection and Segmentation of Abdominal Aortic Aneurysm from Computed Tomography Angiography

Shubhashree Sahoo et al. “A Deep Learning System for the Automated De- tection and Segmentation of Abdominal Aortic Aneurysm from Computed Tomography Angiography”. In:Vascular and Endovascular Review8.16s (2025), pp. 251–259

work page 2025
[5]

A Review of Deep Learning Based Methods for Medical Image Multi-Organ Segmentation

Yabo Fu et al. “A Review of Deep Learning Based Methods for Medical Image Multi-Organ Segmentation”. In:Physics in Medicine & Biology85 (2021), pp. 107–122.doi:10.1016/j.ejmp.2021.05.003

work page doi:10.1016/j.ejmp.2021.05.003 2021
[6]

Ahmedt-Aristizabal, M

PengfeiZhang,YuanzhiCheng,andShinichiTamura.“ShapePrior-Constrained Deep Learning Network for Medical Image Segmentation”. In:Computer- ized Medical Imaging and Graphics(2024).doi: 10.1016/j.compmedimag. 2024.102356

work page doi:10.1016/j.compmedimag 2024
[7]

Attention Gated Networks: Learning to Leverage Salient Regions in Medical Images

Jo Schlemper et al. “Attention Gated Networks: Learning to Leverage Salient Regions in Medical Images”. In:Medical Image Analysis53 (2019), pp. 197–207.doi:10.1016/j.media.2019.01.012

work page doi:10.1016/j.media.2019.01.012 2019
[8]

TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images

Jakob Wasserthal et al. “TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images”. In:Radiology: Artificial Intelligence 5.5 (2023), e230024.doi:10.1148/ryai.230024

work page doi:10.1148/ryai.230024 2023
[9]

The Vascular Modeling Toolkit: A Python Library for the Analysis of Tubular Structures in Medical Images

Richard Izzo et al. “The Vascular Modeling Toolkit: A Python Library for the Analysis of Tubular Structures in Medical Images”. In:Journal of Open Source Software3.25 (2018), p. 745.doi: 10.21105/joss.00745 . url:https://doi.org/10.21105/joss.00745

work page doi:10.21105/joss.00745 2018
[10]

Modeling of size dependent failure in cardio- vascular stent struts under tension and bending

Samarth S. Raut et al. “The Role of Geometric and Biomechanical Factors in Abdominal Aortic Aneurysm Rupture Risk Assessment”. In:Annals of Biomedical Engineering41.7 (2013), pp. 1459–1477.doi:10.1007/s10439- 013- 0786- 6.url: https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC3679219/

work page doi:10.1007/s10439- 2013
[11]

Combined Curvature and Wall Shear Stress Analysis of Abdominal Aortic Aneurysm: An Analysis of Rupture Risk Factors

Biyun Teng et al. “Combined Curvature and Wall Shear Stress Analysis of Abdominal Aortic Aneurysm: An Analysis of Rupture Risk Factors”. In: Cardiovascular and Interventional Radiology45.6 (2022), pp. 752–760.doi: 10.1007/s00270-022-03140-z

work page doi:10.1007/s00270-022-03140-z 2022

[1] [1]

Systematic Review and Meta-Analysis of the Incidence of Rupture, Repair, and Death of Small and Large Abdominal Aortic Aneurysms under Surveillance

Nicola Leone et al. “Systematic Review and Meta-Analysis of the Incidence of Rupture, Repair, and Death of Small and Large Abdominal Aortic Aneurysms under Surveillance”. In:Journal of Clinical Medicine12.21 (2023), p. 6837.doi:10.3390/jcm12216837

work page doi:10.3390/jcm12216837 2023

[2] [2]

Yamagishi, N

Taehun Kim et al. “Computed Tomography-Based Automated Measurement of Abdominal Aortic Aneurysm Using Semantic Segmentation with Active Learning”. In:Scientific Reports14(2024), p. 8924.doi:10.1038/s41598- 024-59735-8

work page doi:10.1038/s41598- 2024

[3] [3]

U-Net: Convolu- tional Networks for Biomedical Image Segmentation

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. “U-Net: Convolu- tional Networks for Biomedical Image Segmentation”. In:Medical Image Computing and Computer-Assisted Intervention (MICCAI). 2015, pp. 234– 241

work page 2015

[4] [4]

A Deep Learning System for the Automated De- tection and Segmentation of Abdominal Aortic Aneurysm from Computed Tomography Angiography

Shubhashree Sahoo et al. “A Deep Learning System for the Automated De- tection and Segmentation of Abdominal Aortic Aneurysm from Computed Tomography Angiography”. In:Vascular and Endovascular Review8.16s (2025), pp. 251–259

work page 2025

[5] [5]

A Review of Deep Learning Based Methods for Medical Image Multi-Organ Segmentation

Yabo Fu et al. “A Review of Deep Learning Based Methods for Medical Image Multi-Organ Segmentation”. In:Physics in Medicine & Biology85 (2021), pp. 107–122.doi:10.1016/j.ejmp.2021.05.003

work page doi:10.1016/j.ejmp.2021.05.003 2021

[6] [6]

Ahmedt-Aristizabal, M

PengfeiZhang,YuanzhiCheng,andShinichiTamura.“ShapePrior-Constrained Deep Learning Network for Medical Image Segmentation”. In:Computer- ized Medical Imaging and Graphics(2024).doi: 10.1016/j.compmedimag. 2024.102356

work page doi:10.1016/j.compmedimag 2024

[7] [7]

Attention Gated Networks: Learning to Leverage Salient Regions in Medical Images

Jo Schlemper et al. “Attention Gated Networks: Learning to Leverage Salient Regions in Medical Images”. In:Medical Image Analysis53 (2019), pp. 197–207.doi:10.1016/j.media.2019.01.012

work page doi:10.1016/j.media.2019.01.012 2019

[8] [8]

TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images

Jakob Wasserthal et al. “TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images”. In:Radiology: Artificial Intelligence 5.5 (2023), e230024.doi:10.1148/ryai.230024

work page doi:10.1148/ryai.230024 2023

[9] [9]

The Vascular Modeling Toolkit: A Python Library for the Analysis of Tubular Structures in Medical Images

Richard Izzo et al. “The Vascular Modeling Toolkit: A Python Library for the Analysis of Tubular Structures in Medical Images”. In:Journal of Open Source Software3.25 (2018), p. 745.doi: 10.21105/joss.00745 . url:https://doi.org/10.21105/joss.00745

work page doi:10.21105/joss.00745 2018

[10] [10]

Modeling of size dependent failure in cardio- vascular stent struts under tension and bending

Samarth S. Raut et al. “The Role of Geometric and Biomechanical Factors in Abdominal Aortic Aneurysm Rupture Risk Assessment”. In:Annals of Biomedical Engineering41.7 (2013), pp. 1459–1477.doi:10.1007/s10439- 013- 0786- 6.url: https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC3679219/

work page doi:10.1007/s10439- 2013

[11] [11]

Combined Curvature and Wall Shear Stress Analysis of Abdominal Aortic Aneurysm: An Analysis of Rupture Risk Factors

Biyun Teng et al. “Combined Curvature and Wall Shear Stress Analysis of Abdominal Aortic Aneurysm: An Analysis of Rupture Risk Factors”. In: Cardiovascular and Interventional Radiology45.6 (2022), pp. 752–760.doi: 10.1007/s00270-022-03140-z

work page doi:10.1007/s00270-022-03140-z 2022