pith. sign in

arxiv: 1907.11392 · v1 · pith:LFIPV4BXnew · submitted 2019-07-26 · 📡 eess.IV · cs.CV

Automatic Calcium Scoring in Cardiac and Chest CT Using DenseRAUnet

Jiechao Ma , Rongguo Zhang This is my paper

Pith reviewed 2026-05-24 15:33 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords coronary artery calcificationCT segmentationDenseRAUnetchest CTcardiac CTcalcium scoringCVD risk predictiondeep learning
0
0 comments X

The pith

DenseRAUnet trained only on chest CT segments coronary calcium on both chest and cardiac CT with F1-score 0.75.

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

The paper aims to enable automatic coronary artery calcium scoring from widely available chest CT scans instead of less common ECG-synchronized cardiac CT. It develops DenseRAUnet, a network blending dense U-Net, ResNet, and atrous convolutions, trained in 2.5D on chest CT only. Testing on a private set of 144 scans covering both types yields an F1-score of 0.75 for segmentation and 0.83 accuracy for risk prediction. This suggests the model generalizes across scan modalities without retraining. A combined bootstrap and IoU loss helps manage the imbalance between calcium pixels and background.

Core claim

DenseRAUnet integrates Dense U-Net, ResNet residual connections, and atrous convolutions into a 2.5D network trained exclusively on chest CT. It segments coronary calcium pixels on both chest and cardiac CT scans from a private dataset of 144 scans, reaching an F1-score of 0.75. The resulting calcium measures support cardiovascular disease risk prediction at 0.83 accuracy. A bootstrap-IoU loss balances the sparse foreground calcium class against background tissue.

What carries the argument

DenseRAUnet, a 2.5D convolutional network that fuses dense skip connections, residual blocks, and atrous convolutions, trained with a bootstrap-enhanced IoU loss to segment sparse calcium regions.

If this is right

  • Routine chest CT scans can be used for CAC quantification and CVD risk assessment.
  • No need for separate models or ECG-gated cardiac CT for calcium scoring.
  • The architecture allows generalization from chest to cardiac CT without fine-tuning.
  • Accurate segmentation supports reliable risk prediction at 0.83 level.

Where Pith is reading between the lines

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

  • Extending the approach to multi-vendor public datasets could validate broader clinical deployment.
  • The loss function design may apply to other medical segmentation tasks with severe class imbalance.
  • 2.5D processing could reduce computational demands for similar 3D medical imaging problems.

Load-bearing premise

The private dataset of 144 scans captures the full range of calcium appearances, patient demographics, and scanner characteristics found in real-world clinical use.

What would settle it

Evaluating the model on a larger, publicly available dataset of chest and cardiac CT scans acquired from different scanners and patient groups would reveal if the reported F1-score and accuracy hold.

Figures

Figures reproduced from arXiv: 1907.11392 by Jiechao Ma, Rongguo Zhang.

Figure 1
Figure 1. Figure 1: The overall structure of DenseUnet and details in Residual Atrous Unit The basic network is an encoder-decoder architecture, similar to dense U-Net. We adopt a backbone network (DenseNet-121) as the encoder sub-network. The decoder sub-network consists of three decoder modules. Each decoder module is [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Segmentation results of chest CT (top) and cardiac CT (bottom). From left to right: the segmentation result of our model without post-processing, the result of with post-processing, ground truth. References 1. John A Rumberger, Bruce H Brundage, Daniel J Rader, and George Kondos. Electron beam computed tomographic coronary calcium scanning: a review and guidelines for use in asymptomatic persons. In Mayo C… view at source ↗
read the original abstract

Cardiovascular disease (CVD) is a common and strong threat to human beings, featuring high prevalence, disability and mortality. The amount of coronary artery calcification (CAC) is an effective factor for CVD risk evaluation. Conventionally, CAC is quantified using ECG-synchronized cardiac CT but rarely from general chest CT scans. However, compared with ECG-synchronized cardiac CT, chest CT is more prevalent and economical in clinical practice. To address this, we propose an automatic method based on Dense U-Net to segment coronary calcium pixels on both types of CT scans. Our contribution is two-fold. First, we propose a novel network called DenseRAUnet, which takes advantage of Dense U-net, ResNet and atrous convolutions. We prove the robustness and generalizability of our model by training it exclusively on chest CT while test on both types of CT scans. Second, we design a loss function combining bootstrap with IoU function to balance foreground and background classes. DenseRAUnet is trained in a 2.5D fashion and tested on a private dataset consisting of 144 scans. Results show an F1-score of 0.75, with 0.83 accuracy of predicting cardiovascular disease risk.

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

3 major / 2 minor

Summary. The paper proposes DenseRAUnet, a 2.5D network combining Dense U-Net, ResNet blocks, and atrous convolutions, for segmenting coronary artery calcium on CT. Trained exclusively on chest CT scans using a bootstrap-IoU loss, the model is evaluated on a private dataset of 144 scans (both chest and cardiac CT) and reports an F1-score of 0.75 for calcium segmentation together with 0.83 accuracy for CVD risk prediction.

Significance. If the reported performance holds under proper validation, the work would be significant because it targets CAC quantification from the far more prevalent and economical chest CT rather than ECG-gated cardiac CT. The architectural combination and class-imbalance loss are sensible design choices that could be reusable; however, the absence of any public benchmark, external validation, or baseline comparison prevents assessment of whether the numbers advance the state of the art.

major comments (3)
  1. [Abstract] Abstract: the central performance claims (F1 = 0.75, risk-prediction accuracy = 0.83) are presented without any baseline method, without cross-validation procedure, without error bars or confidence intervals, and without stating how the 144 scans were partitioned or selected. These omissions make the numerical results impossible to interpret or reproduce and directly undermine the generalizability assertion.
  2. [Methods / Dataset] Dataset description (Methods): no information is supplied on scanner vendors, reconstruction kernels, slice thickness, patient demographics, or the proportion of chest versus cardiac cases within the 144-scan private set. Because the central claim is that training on chest CT alone yields robust performance on both modalities, the lack of acquisition-parameter diversity documentation is load-bearing.
  3. [Results / Evaluation] Evaluation protocol: the paper states that the model was “tested on both types of CT scans” yet provides neither a per-modality breakdown of the F1 score nor any statistical test comparing chest-only versus mixed test performance. Without this disaggregation the robustness claim cannot be verified.
minor comments (2)
  1. [Abstract / Introduction] The abstract and introduction contain several minor grammatical issues and inconsistent capitalization of “DenseRAUnet” versus “Dense U-Net.”
  2. [Figures] Figure captions and axis labels should explicitly state the number of test cases and whether the displayed examples are from chest or cardiac CT.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below and indicate the revisions planned for the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claims (F1 = 0.75, risk-prediction accuracy = 0.83) are presented without any baseline method, without cross-validation procedure, without error bars or confidence intervals, and without stating how the 144 scans were partitioned or selected. These omissions make the numerical results impossible to interpret or reproduce and directly undermine the generalizability assertion.

    Authors: We agree additional context is needed for interpretability. In revision we will expand the abstract to note that the model was trained exclusively on chest CT and evaluated on a private set of 144 scans (both modalities) using a single random train-test split. Bootstrap-derived confidence intervals will be added to the reported metrics in the results. Cross-validation was omitted to maximize training data given the limited private set size. No direct baseline comparisons were performed in the original submission; we will add discussion of related methods but note that public implementations for this exact task were unavailable at submission time. revision: partial

  2. Referee: [Methods / Dataset] Dataset description (Methods): no information is supplied on scanner vendors, reconstruction kernels, slice thickness, patient demographics, or the proportion of chest versus cardiac cases within the 144-scan private set. Because the central claim is that training on chest CT alone yields robust performance on both modalities, the lack of acquisition-parameter diversity documentation is load-bearing.

    Authors: We will revise the Methods section to document all recorded acquisition parameters (scanner vendors, kernels, slice thickness), patient demographics, and the precise proportion of chest versus cardiac cases within the 144-scan private set. This will directly support the cross-modality robustness claim. revision: yes

  3. Referee: [Results / Evaluation] Evaluation protocol: the paper states that the model was “tested on both types of CT scans” yet provides neither a per-modality breakdown of the F1 score nor any statistical test comparing chest-only versus mixed test performance. Without this disaggregation the robustness claim cannot be verified.

    Authors: We will add a per-modality F1-score breakdown in the revised Results section. Statistical comparisons (e.g., appropriate tests for the subset sizes) between chest-CT and cardiac-CT performance will also be reported to allow verification of the robustness claim. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML metrics on held-out private data with no derivations

full rationale

The paper reports direct empirical results from training DenseRAUnet on chest CT and evaluating F1-score and risk-prediction accuracy on a private 144-scan test set. No equations, parameter-fitting steps, self-citations, or derivations are described that could reduce claims to inputs by construction. The generalizability statement is an empirical claim about the chosen dataset split, not a mathematical reduction. This is the standard non-circular outcome for a methods/results paper without theoretical derivations.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central performance numbers rest on the empirical behavior of a neural network whose weights are learned from the private training set, plus the domain assumption that 2.5D slices suffice for 3D context and that the bootstrap-IoU loss balances classes without introducing bias.

free parameters (2)
  • network weights and hyperparameters
    All convolutional filter values and training hyperparameters are fitted to the chest CT training data.
  • bootstrap sampling ratio in loss
    The mixing weight between bootstrap and IoU terms is chosen to balance foreground/background but is not derived from first principles.
axioms (1)
  • domain assumption 2.5D slice stacking captures adequate 3D spatial context for accurate calcium segmentation
    Invoked by the choice of 2.5D training mode without further justification in the abstract.

pith-pipeline@v0.9.0 · 5743 in / 1294 out tokens · 25533 ms · 2026-05-24T15:33:59.043752+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

14 extracted references · 14 canonical work pages

  1. [1]

    Electron beam computed tomographic coronary calcium scanning: a review and guidelines for use in asymptomatic persons

    John A Rumberger, Bruce H Brundage, Daniel J Rader, and George Kondos. Electron beam computed tomographic coronary calcium scanning: a review and guidelines for use in asymptomatic persons. In Mayo Clinic Proceedings, volume 74, pages 243–252. Elsevier, 1999

    work page 1999

  2. [2]

    Growing a random forest with fuzzy spatial features for fully automatic artery-specific coronary calcium scoring

    Felix Durlak, Michael Wels, Chris Schwemmer, Michael S¨ uhling, Stefan Steidl, and Andreas Maier. Growing a random forest with fuzzy spatial features for fully automatic artery-specific coronary calcium scoring. In International Workshop on Machine Learning in Medical Imaging , pages 27–35. Springer, 2017

    work page 2017

  3. [3]

    Automatic coronary calcium scoring in low-dose chest computed tomography

    Ivana Isgum, Mathias Prokop, Meindert Niemeijer, Max A Viergever, and Bram Van Ginneken. Automatic coronary calcium scoring in low-dose chest computed tomography. IEEE transactions on medical imaging , 31(12):2322–2334, 2012. 8 Jiechao Ma and Rongguo Zhang

    work page 2012

  4. [4]

    A supervised classification-based method for coronary calcium de- tection in non-contrast ct

    Uday Kurkure, Deepak R Chittajallu, Gerd Brunner, Yen H Le, and Ioannis A Kakadiaris. A supervised classification-based method for coronary calcium de- tection in non-contrast ct. The international journal of cardiovascular imaging , 26(7):817–828, 2010

    work page 2010

  5. [5]

    Vessel specific coronary artery calcium scoring: an automatic system

    Rahil Shahzad, Theo van Walsum, Michiel Schaap, Alexia Rossi, Stefan Klein, Annick C Weustink, Pim J de Feyter, Lucas J van Vliet, and Wiro J Niessen. Vessel specific coronary artery calcium scoring: an automatic system. Academic radiology, 20(1):1–9, 2013

    work page 2013

  6. [6]

    An automatic machine learning system for coronary calcium scoring in clinical non-contrast enhanced, ecg-triggered cardiac ct

    Jelmer M Wolterink, Tim Leiner, Richard AP Takx, Max A Viergever, and Ivana Iˇ sgum. An automatic machine learning system for coronary calcium scoring in clinical non-contrast enhanced, ecg-triggered cardiac ct. In Medical Imaging 2014: Computer-Aided Diagnosis , volume 9035, page 90350E. International Society for Optics and Photonics, 2014

    work page 2014

  7. [7]

    Automatic coronary calcium scoring in cardiac ct angiography using convolutional neural net- works

    Jelmer M Wolterink, Tim Leiner, Max A Viergever, and Ivana Iˇ sgum. Automatic coronary calcium scoring in cardiac ct angiography using convolutional neural net- works. In International Conference on Medical Image Computing and Computer- Assisted Intervention, pages 589–596. Springer, 2015

    work page 2015

  8. [8]

    Automatic calcium scoring in low-dose chest ct using deep neural networks with dilated convolutions

    Nikolas Lessmann, Bram van Ginneken, Majd Zreik, Pim A de Jong, Bob D de Vos, Max A Viergever, and Ivana Iˇ sgum. Automatic calcium scoring in low-dose chest ct using deep neural networks with dilated convolutions. IEEE transactions on medical imaging, 37(2):615–625, 2018

    work page 2018

  9. [9]

    Automatic coronary artery calcium scoring in cardiac ct angiography using paired convolutional neural networks

    Jelmer M Wolterink, Tim Leiner, Bob D de Vos, Robbert W van Hamersvelt, Max A Viergever, and Ivana Iˇ sgum. Automatic coronary artery calcium scoring in cardiac ct angiography using paired convolutional neural networks. Medical image analysis, 34:123–136, 2016

    work page 2016

  10. [10]

    Fully- convolutional deep-learning based system for coronary calcium score prediction from non-contrast chest ct

    Ran Shadmi, Victoria Mazo, Orna Bregman-Amitai, and Eldad Elnekave. Fully- convolutional deep-learning based system for coronary calcium score prediction from non-contrast chest ct. In 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018) , pages 24–28. IEEE, 2018

    work page 2018

  11. [11]

    Densely connected convolutional networks

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 4700–4708, 2017

    work page 2017

  12. [12]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence , 40(4):834–848, 2018

    work page 2018

  13. [13]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1–9, 2015

    work page 2015

  14. [14]

    Concurrent spatial and channel squeeze & excitationin fully convolutional networks

    Abhijit Guha Roy, Nassir Navab, and Christian Wachinger. Concurrent spatial and channel squeeze & excitationin fully convolutional networks. In International Con- ference on Medical Image Computing and Computer-Assisted Intervention , pages 421–429. Springer, 2018

    work page 2018