pith. sign in

arxiv: 2607.00633 · v1 · pith:3K4H7OGPnew · submitted 2026-07-01 · 💻 cs.CE

Stacked Ensemble Learning for Abdominal Aortic Aneurysm Segmentation in CT Angiography

Joshua Fry , Sajjad Arzemanzadeh , Saeideh Sekhavat , Mostafa Jamshidian , Adam Wittek , Michael Bertolacci , Elke R. Gizewski , Eva Gassner
show 4 more authors
Alexander Loizides Maximillian Lutz Florian K. Enzmann Karol Miller
This is my paper

Pith reviewed 2026-07-02 03:32 UTC · model grok-4.3

classification 💻 cs.CE
keywords abdominal aortic aneurysmCTA segmentationstacked ensemble learningnnUNetDice coefficientSeparation Distance
0
0 comments X

The pith

A stacked ensemble of three nnUNetv2 models outperforms the base learners on AAA segmentation from CTA, reaching a mean Dice of 0.9752 and mean Separation Distance of 0.4598 mm on held-out cases.

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

The paper develops a stacked ensemble framework to automate segmentation of abdominal aortic aneurysms from computed tomography angiography images. It trains three variants of nnUNetv2 as base learners and combines their probability outputs with a logistic regression meta-model. On eight held-out test cases, the ensemble outperforms the individual models in both volumetric overlap and boundary accuracy. This matters because accurate 3D aneurysm geometry is required for patient-specific biomechanical computations that assess rupture risk, and manual segmentation is time-consuming and observer-dependent.

Core claim

The stacked ensemble learning approach for AAA segmentation from CTA images, which combines voxel-wise probability outputs from three nnUNetv2 configurations using an L2-regularised logistic regression meta-model trained on out-of-sample cross-validation predictions, produces improved segmentation performance on the held-out test set, with a mean Dice Coefficient of 0.9752 and a mean Separation Distance of 0.4598 mm, compared to the base learners.

What carries the argument

The L2-regularised logistic regression meta-model that combines probability outputs from three nnUNetv2 base learners trained on out-of-sample cross-validation predictions.

If this is right

  • The ensemble improves volumetric overlap and average boundary agreement relative to the three nnUNetv2 base learners.
  • Boundary accuracy gains support more reliable patient-specific biomechanical computations for rupture risk assessment.
  • The introduced Separation Distance metric quantifies mean surface-to-surface agreement between segmentations.

Where Pith is reading between the lines

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

  • The framework could reduce the time and observer variability that currently limit large-scale AAA studies.
  • Testing the same stacking procedure on other vascular or organ segmentation tasks would be a direct next step.
  • Multi-center validation would be required before the method could be considered ready for routine clinical deployment.

Load-bearing premise

The eight held-out test cases are representative of the variability encountered in future clinical CTA scans.

What would settle it

Evaluating the ensemble on a larger multi-center collection of CTA scans in which it fails to outperform the individual base learners would falsify the central performance claim.

read the original abstract

Abdominal aortic aneurysm (AAA) rupture risk assessment increasingly relies on patient-specific biomechanical computations, which require accurate three-dimensional aneurysm geometry from computed tomography angiography (CTA). Manual and semi-automated segmentation remain time-consuming and observer-dependent, limiting their use in large-scale clinical workflows. In this study, we developed a stacked ensemble framework for automated AAA seg-mentation from CTA images. We used 40 anonymised contrast-enhanced CTA scans from AAA patients and generated reference segmentations using the nnInteractive extension in 3D Slicer. We partitioned the dataset into 32 training cases and 8 held-out test cases. Three nnUNetv2 configurations, Default, DA5, and ResEncL, were trained as base learners, and their voxel-wise probability out-puts were combined using an L2-regularised logistic regression meta-model trained from out-of-sample cross-validation predictions. We evaluated segmentation performance using Dice Coefficient and Separation Distance, a mean boundary-to-boundary distance measure introduced in this study to quantify average surface agreement. On the held-out test set, the ensemble achieved the highest mean Dice Coefficient of 0.9752 and the lowest mean Separation Distance of 0.4598 mm, indicating improved volumetric overlap and average boundary agreement compared with the individual base learners. Overall, stacked ensemble learning provided small but meaningful improvements in AAA segmentation, particularly for boundary accuracy relevant to downstream patient-specific bio-mechanical computations.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript describes a stacked ensemble framework for AAA segmentation in CTA images. Three nnUNetv2 base learners (Default, DA5, ResEncL) are trained on 32 cases; their voxel-wise probability outputs are combined via an L2-regularized logistic regression meta-learner trained on out-of-fold predictions from cross-validation. On an 8-case held-out test set the ensemble is reported to achieve the highest mean Dice Coefficient (0.9752) and lowest mean Separation Distance (0.4598 mm) relative to the individual base learners.

Significance. If the reported gains prove robust, the work offers a practical incremental improvement in automated AAA segmentation accuracy, particularly for boundary fidelity needed in downstream patient-specific biomechanical modeling. The empirical results on held-out data constitute a clear, falsifiable claim; however, the small test-set size and lack of statistical support limit the strength of any generalizability assertion.

major comments (2)
  1. [Results / held-out test set evaluation] Test-set evaluation (n=8): superiority is asserted solely from aggregate means (Dice 0.9752, Separation Distance 0.4598 mm) with no per-case scores, standard deviations, confidence intervals, or paired statistical tests (Wilcoxon or t-test on differences) reported. This directly undermines the central claim that the ensemble provides meaningful improvement over the base learners.
  2. [Methods / metric definition] Separation Distance metric: the new boundary-to-boundary distance is introduced and used to support the claim of 'improved average boundary agreement' without any comparison to established surface-distance measures (mean surface distance, 95 % Hausdorff) or external validation demonstrating added clinical or biomechanical utility.
minor comments (2)
  1. [Abstract] Abstract contains a hyphenation artifact ('seg-mentation').
  2. [Methods] The L2 regularization strength of the meta-learner is a free hyper-parameter; the manuscript should state how its value was selected and whether any sensitivity analysis was performed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major point below, agreeing where the critique is valid and outlining specific revisions to strengthen the presentation of results and metric evaluation.

read point-by-point responses

  1. Referee: [Results / held-out test set evaluation] Test-set evaluation (n=8): superiority is asserted solely from aggregate means (Dice 0.9752, Separation Distance 0.4598 mm) with no per-case scores, standard deviations, confidence intervals, or paired statistical tests (Wilcoxon or t-test on differences) reported. This directly undermines the central claim that the ensemble provides meaningful improvement over the base learners.

    Authors: We agree that aggregate means alone are insufficient to support claims of improvement. In the revision we will add a supplementary table listing per-case Dice and Separation Distance values for the ensemble and all three base learners on the eight test cases. We will also report standard deviations and 95% confidence intervals around the means. Paired Wilcoxon signed-rank tests will be performed on the per-case differences and p-values reported. Because the test set is small, we will explicitly note the limited statistical power and frame the comparisons as descriptive rather than definitive evidence of superiority. revision: yes

  2. Referee: [Methods / metric definition] Separation Distance metric: the new boundary-to-boundary distance is introduced and used to support the claim of 'improved average boundary agreement' without any comparison to established surface-distance measures (mean surface distance, 95 % Hausdorff) or external validation demonstrating added clinical or biomechanical utility.

    Authors: The Separation Distance was introduced to provide an intuitive average boundary-to-boundary measure relevant to downstream biomechanical modeling. We accept that direct comparisons to standard surface metrics are required. The revised manuscript will include mean surface distance and 95% Hausdorff distance computed on the same test set for all methods, presented alongside the new metric. We will also expand the discussion to explain the metric's intended utility for wall-stress calculations while acknowledging that a dedicated external clinical validation study lies outside the current scope. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML evaluation on held-out data

full rationale

The paper reports standard supervised learning results: three nnUNetv2 base models are trained on 32 cases, their out-of-fold predictions train an L2-regularised logistic regression meta-learner, and all models are evaluated with Dice and a newly introduced Separation Distance metric on an independent 8-case held-out test set. No equations, uniqueness theorems, or first-principles derivations are present; the performance numbers (0.9752 Dice, 0.4598 mm Separation Distance) are direct empirical measurements, not quantities forced by construction from the training procedure itself. The new metric is defined once and applied uniformly; it does not create a self-referential loop. Self-citation load-bearing, ansatz smuggling, or renaming of known results are absent. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

Only the abstract is available; the ledger is therefore limited to what can be read from it. The central claim rests on the assumption that nnInteractive reference labels are sufficiently accurate ground truth and that the 8-case test set captures relevant variability.

free parameters (1)
  • L2 regularization strength in logistic regression meta-learner
    Chosen during training; value not stated in abstract.
axioms (2)
  • domain assumption nnInteractive labels in 3D Slicer constitute reliable reference segmentations
    Invoked when the 40 scans are labeled and used as training targets.
  • domain assumption The 8 held-out cases are statistically representative of future clinical CTA variability
    Required for any claim that the reported metrics will generalize.

pith-pipeline@v0.9.1-grok · 5836 in / 1460 out tokens · 37628 ms · 2026-07-02T03:32:12.523449+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 2 canonical work pages

  1. [1]

    Eur J Vasc Endovasc Surg 57, 8–93 (2019)

    Wanhainen, A., Verzini, F., Van Herzeele, I., Allaire, E., Bown, M., Cohnert, T., Dick, F., van Herwaarden, J., Karkos, C., Koelemay, M., Kölbel, T., Loftus, I., Mani, K., Melissano, G., Powell, J., Szeberin, Z., Esvs Guidelines, C., de Borst, G.J., Chakfe, N., Debus, S., Hinchliffe, R., Kakkos, S., Koncar, I., Kolh, P., Lindholt, J.S., de Vega, M., Verma...

    2019

  2. [2]

    European Journal of Vascular and Endovascular Surgery 17, 472–475 (1999)

    Lindholt, J.S., Vammen, S., Juul, S., Henneberg, E.W., Fasting, H.: The Validity of Ultrasonographic Scanning as Screening Method for Abdominal Aortic Aneurysm. European Journal of Vascular and Endovascular Surgery 17, 472–475 (1999)

    1999

  3. [3]

    British Journal of Surgery 103, 1634–1639 (2016)

    Laine, M.T., Laukontaus, S.J., Kantonen, I., Venermo, M.: Population -based study of ruptured abdominal aortic aneurysm. British Journal of Surgery 103, 1634–1639 (2016)

    2016

  4. [4]

    NICE: Abdominal aortic aneurysm: diagnosis and management. (2020)

    2020

  5. [5]

    The case for early resection

    Darling, R.C., Messina, C.R., Brewster, D.C., Ottinger, L.W.: Autopsy study of unoperated abdominal aortic aneurysms. The case for early resection. Circulation 56, II161–164 (1977)

    1977

  6. [6]

    Ann R Coll Surg Engl 81, 27–31 (1999)

    Choksy, S.A., Wilmink, A.B., Quick, C.R.: Ruptured abdominal aortic aneurysm in the Huntingdon district: a 10-year experience. Ann R Coll Surg Engl 81, 27–31 (1999)

    1999

  7. [7]

    Front Surg 3, 1 (2016)

    Kontopodis, N., Pantidis, D., Dedes, A., Daskalakis, N., Ioannou, C.V.: The - Not So - Solid 5.5 cm Threshold for Abdominal Aortic Aneurysm Repair: Facts, Misinterpretations, and Future Directions. Front Surg 3, 1 (2016)

    2016

  8. [8]

    J Vasc Surg 37, 724–732 (2003)

    Fillinger, M.F., Marra, S.P., Raghavan, M.L., Kennedy, F.E.: Prediction of Rupture Risk in Abdominal Aortic Aneurysm during Observation: Wall Stress versus Diameter. J Vasc Surg 37, 724–732 (2003)

    2003

  9. [9]

    European Journal of Vascular and Endovascular Surgery 40, 176–185 (2010)

    Gasser, T.C., Auer, M., Labruto, F., Swedenborg, J., Roy, J.: Biomechanical Rupture Risk Assessment of Abdominal Aortic Aneurysms: Model Complexity versus Predictability of Finite Element Simulations. European Journal of Vascular and Endovascular Surgery 40, 176–185 (2010)

    2010

  10. [10]

    Scientific Reports 7, (2017)

    Joldes, G.R., Miller, K., Wittek, A., Forsythe, R.O., Newby, D.E., Doyle, B.J.: BioPARR: A software system for estimating the rupture potential index for abdominal aortic aneurysms. Scientific Reports 7, (2017)

    2017

  11. [11]

    J Vasc Surg 71, 617–+ (2020)

    Polzer, S., Gasser, T.C., Vlachovsky, R., Kubícek, L., Lambert, L., Man, V., Novák, K., Slazansky, M., Bursa, J., Staffa, R.: Biomechanical indices are more sensitive than diameter in predicting rupture of asymptomatic abdominal aortic aneurysms. J Vasc Surg 71, 617–+ (2020)

    2020

  12. [12]

    J Mech Behav Biomed 58, 139–148 (2016)

    Joldes, G.R., Miller, K., Wittek, A., Doyle, B.: A simple, effective and clinically applicable method to compute abdominal aortic aneurysm wall stress. J Mech Behav Biomed 58, 139–148 (2016)

    2016

  13. [13]

    Int J Numer Meth Bio 38, (2022)

    Wittek, A., Alkhatib, F., Vitasek, R., Polzer, S., Miller, K.: On stress in abdominal aortic aneurysm: Linear versus non-linear analysis and aneurysms rupture risk. Int J Numer Meth Bio 38, (2022)

    2022

  14. [14]

    Lecture Notes in Bioengineering, pp

    Jamshidian, M., Sekhavat, S., Wittek, A., Le Liepvre, D., Bernard, F., Minvielle, L., Fondanèche, A., Miller, K.: Abdominal aortic aneurysm wall stress: A 7-line code in MATLAB 13 and a one -click software application. Lecture Notes in Bioengineering, pp. 88 –103. Springer (2024)

    2024

  15. [15]

    J Biomech 179, 112484 (2025)

    Jamshidian, M., Wittek, A., Sekhavat, S., Miller, K.: Kinematics of abdominal aortic Aneurysms. J Biomech 179, 112484 (2025)

    2025

  16. [16]

    J Ultras Med 42, 1737–1746 (2023)

    Derwich, W., Keller, T., Filmann, N., Schmitz-Rixen, T., Blasé, C., Oikonomou, K., Wittek, A.: Changes in Aortic Diameter and Wall Strain in Progressing Abdominal Aortic Aneurysms. J Ultras Med 42, 1737–1746 (2023)

    2023

  17. [17]

    J Biomech 48, 354–360 (2015)

    Satriano, A., Rivolo, S., Martufi, G., Finol, E.A., Di Martino, E.S.: In vivo strain assessment of the abdominal aortic aneurysm. J Biomech 48, 354–360 (2015)

    2015

  18. [18]

    Int J Numer Meth Bio 42, (2026)

    Jamshidian, M., Wittek, A., Sekhavat, S., Mufty, H., Maleux, G., Fourneau, I., Gizewski, E.R., Gassner, E., Loizides, A., Lutz, M., Enzmann, F.K., Le Liepvre, D., Bernard, F., Minvielle, L., Fondanèche, A., Miller, K.: Towards Personalised Assessment of Abdominal Aortic Aneurysm Structural Integrity. Int J Numer Meth Bio 42, (2026)

    2026

  19. [19]

    Acta Biomaterialia (2026)

    Schönborn, M., Hegner, A., Derwich, W., Oikonomou, K., Huß, A., Gámez, A.J., Wittek, A., Blase, C.: An efficient framework to assess the predictive value of 4D ultrasound -based wall motion indices for abdominal aortic aneurysm rupture risk stratification. Acta Biomaterialia (2026)

    2026

  20. [20]

    arXiv preprint arXiv:2509.12550 (2025)

    Sekhavat, S., Jamshidian, M., Wittek, A., Miller, K.: Impact of Geometric Uncertainty on the Computation of Abdominal Aortic Aneurysm Wall Strain. arXiv preprint arXiv:2509.12550 (2025)

    work page arXiv 2025

  21. [21]

    In: Nash, M.P., Wittek, A., Nielsen, P.M.F., Kobielarz, M., Babu, A.R., Miller, K

    Hodge, T., Tan, J.C.Y., Koh, P.H., Storer, E., Huynh, A., Alkhatib, F., Miller, K., Wittek, A.: Effect of Analyst Segmentation Variability on Computed AAA Stress Distributions. In: Nash, M.P., Wittek, A., Nielsen, P.M.F., Kobielarz, M., Babu, A.R., Miller, K. (eds.) Lecture Notes in Bioengineering, pp. 63–77. Springer (2023)

    2023

  22. [22]

    Lecture Notes in Bioengineering, pp

    Gralton, S.G., Alkhatib, F., Zwick, B., Bourantas, G., Wittek, A., Miller, K.: Random boundaries: quantifying segmentation uncertainty in solutions to boundary- value problems. Lecture Notes in Bioengineering, pp. 17–32. Springer (2023)

    2023

  23. [23]

    Computers in Biology and Medicine 190, 110017 (2025)

    Bošnjak, D., Schussnig, R., Ranftl, S., Holzapfel, G.A., Fries, T.- P.: Geometric uncertainty of patient -specific blood vessels and its impact on aortic hemodynamics: A computational study. Computers in Biology and Medicine 190, 110017 (2025)

    2025

  24. [24]

    arXiv (2024)

    Alkhatib, F., Jamshidian, M., Le Liepvre, D., Bernard , F., Minvielle, L., Wittek, A., Miller, K.: Towards Full Automation of Geometry Extraction for Biomechanical Analysis of Abdominal Aortic Aneurysm; Neural Network-Based versus Classical Methodologies. arXiv (2024)

    2024

  25. [25]

    Nature Methods 18, 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier -Hein, K.H.: nnU-Net: a self - configuring method for deep learning-based biomedical image segmentation. Nature Methods 18, 203–211 (2021)

    2021

  26. [26]

    Krizhevsky, A.S., I; Hinton, G: ImageNet Classification with Deep Convolutional Neural Networks. (2012)

    2012

  27. [27]

    Ronneberger, O., Fischer, P., Brox, T.: U -Net: Convolutional Networks for Biomedical Image Segmentation. pp. 234–241. Springer International Publishing, Cham (2015)

    2015

  28. [28]

    Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier -Hein, K., Jäger, P.F.: nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation. pp. 488–498. Springer Nature Switzerland, Cham (2024) 14

    2024

  29. [29]

    Journal of Medical Imaging and Radiation Oncology 65, 545–563 (2021)

    Chlap, P., Min, H., Vandenberg, N., Dowling, J., Holloway, L., Haworth, A.: A review of medical image data augmentation techniques for deep learning applications. Journal of Medical Imaging and Radiation Oncology 65, 545–563 (2021)

    2021

  30. [30]

    Imran, M., Krebs, J.R., Balaji Sivaraman, V., Zhang, T., Kumar, A., Ueland, W.R., Fassler, M.J., Huang, J., Sun, X., Wang, L., Shi, P., Rokuss, M., Baumgartner, M., Kirchhof, Y., Maier - Hein, K.H., Isensee, F., Liu, S., Han, B., Thanh Nguyen, B., Shin, D.-j., Ji -Woo, P., Choi, M., Uhm, K.-H., Ko, S.-J., Lee, C., Chun, J., Kim, J.S., Zhang, M., Zhang, H....

    2025

  31. [31]

    IEEE Access 10, 99129–99149 (2022)

    Mienye, I.D., Sun, Y.: A Survey of Ensemble Learning: Concepts, Algorithms, Applications, and Prospects. IEEE Access 10, 99129–99149 (2022)

    2022

  32. [32]

    2018 IEEE Second International Conference on Data Stream Mining & Processing (DSMP), pp

    Pavlyshenko, B.: Using Stacking Approaches for Machine Learning Models. 2018 IEEE Second International Conference on Data Stream Mining & Processing (DSMP), pp. 255 –258 (2018)

    2018

  33. [33]

    Gupta, N., Smith, J., Adlam, B., Mariet, Z.: Ensembling over Classifiers: a Bias -Variance Perspective (2022)

    2022

  34. [34]

    GitHub (2024)

    MaxwellEng: MICCAI_CHANLLENGE24_HJL. GitHub (2024)

    2024

  35. [35]

    de Vente, C., Vaidhya Venkadesh, K., van Ginneken, B., Sánchez, C.I.: SlicerNNInteractive: A 3D Slicer extension for nnInteractive. pp. arXiv:2504.07991 (2025)

    work page arXiv 2025

  36. [36]

    Magn Reson Imaging 30, 1323–1341 (2012)

    Fedorov, A., Beichel, R., Kalpathy-Cramer, J., Finet, J., Fillion-Robin, J.C., Pujol, S., Bauer, C., Jennings, D., Fennessy, F., Sonka, M., Buatti, J., Aylward, S., Miller, J.V., Pieper, S., Kikinis, R.: 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging 30, 1323–1341 (2012)

    2012

  37. [37]

    GitHub (2025)

    Isensee, F.: nnUNet. GitHub (2025)

    2025

  38. [38]

    Springer (2017)

    Hastie, T., Friedman, J., Tibshirani, R.: The Elements of Statistical Learning. Springer (2017)

    2017

  39. [39]

    Ecology 26, 297–302 (1945)

    Dice, L.R.: Measures of the Amount of Ecologic Association Between Species. Ecology 26, 297–302 (1945)

    1945