Sage Journals: Discover world-class research

Abstract

Background

Calcification of abdominal arteries is an important risk marker in vascular disease. Automated, objective quantification methods could improve reproducibility and reduce observer dependency in clinical practice.

Purpose

To develop and evaluate a deep learning method for quantifying abdominal arterial calcification from contrast-enhanced CT angiography (CTA).

Material and Methods

We retrospectively collected 223 CTA volumes, divided into 147 training and 76 test cases. Ground truth calcification segmentations were manually annotated, while vessel segmentations were generated by a previously trained neural network and manually refined. Two nnU-Net models were trained, one for artery segmentation and one for calcification segmentation. Renal, mesenteric, and common iliac arteries were shortened algorithmically. Performance of the models was evaluated using Dice score, volumetric similarity, sensitivity, precision, and Jaccard index. Calcification burden was defined as the ratio of calcified volume to artery volume. The amount and the average size of calcification clusters were investigated. The performance of the method was benchmarked against an idealized threshold-based approach and a more clinically realistic approach.

Results

The neural network achieved performance comparable to the optimized threshold-based method, with slight improvements across several segmentation metrics. Dice scores and volumetric similarity demonstrated reliable vessel and calcification detection. The predicted calcification burden score showed high correlation with the ground truth calcification burden score.

Conclusion

The proposed deep learning tool enables fast, reproducible, and observer-independent quantification of calcification in major abdominal vessels, offering a practical alternative to manual or threshold-based scoring methods.

Keywords

adults arteries arteriosclerosis computer applications-detection diagnosis CT-angiography modalities/techniques structures subject matter topics

Introduction

Peripheral artery disease (PAD) affects 200 million people globally and is a major risk factor for cardiovascular mortality.^1–3 Projections suggest a 50% increase in prevalence by 2045, with a concerning rise among younger populations.^4–7 PAD is associated with a substantial burden of cardiovascular comorbidities and risk factors, reflecting the nature of systemic atherosclerosis.⁸ It is characterized by progressive atherosclerotic plaques within the distal aorta, iliac, and lower limb arteries. However, atherosclerosis and specific calcification of the renal and mesenteric arteries remain less well-investigated areas.

The effect of calcification of the different abdominal arteries has been widely studied and has been found to have a correlation with different comorbidities and surgical complications.^9–24 These evaluations are commonly performed on non-contrast CT scans using the Agatston method, a well-established scoring system. The evaluation of the severity of the calcification of the arteries automatically has been done from unenhanced CT images.²⁵ However, there has been an increased interest in determining calcification from contrast-enhanced CT (CTA) images.^26–28 Since CTA is already widely used in vascular imaging, integrating calcification scoring into existing workflows could eliminate the need for additional non-contrast scans for the purpose of calcification scoring, thereby reducing patient radiation exposure.

Despite these advantages, quantifying calcification from CTA presents technical challenges. Conventional threshold-based algorithms, which are commonly used in non-enhanced scans, perform worse for CTA due to the contrast agent’s impact on Hounsfield Unit (HU) variability.^29,30

In this study, we present a fully automated neural-network–based system for quantifying vascular calcification from CTA. Our method utilizes two neural network segmentation models; one network identifies the arteries while the second identifies the calcifications. This is supported by a post-processing algorithm that standardizes the branching lengths of the abdominal aorta and its side arteries. The neural network separates the abdominal arterial system into four clinically meaningful groups: the aorta, renal arteries, mesenteric arteries, and iliac arteries. By implementing a volume-normalized scoring framework, our system accounts for both anatomical variability and artery segmentation differences. Finally, we validate this model by comparing its performance against both idealized and clinically realistic automated thresholding methods.

The aim of this study is to establish a fast, reproducible, and observer-independent neural-network based algorithm for the calcification scoring of the abdominal aorta and its major arteries from CTA, providing a robust alternative to traditional threshold-based tools for the abdominal aorta and its major branches.

Materials and methods

Data

The study was authorized by the institutional review board for research (HUS Diagnostic Center decisions HUS/211/2020, 23 March 2020). The study was deemed exempt for the requirement for signed informed consent according to the Finnish research law (488/1999 and 552/2019).

A total of 223 CTA volumes, acquired between 2011 and 2013 using standard clinical protocols at Turku University Hospital, were included. The volumes were divided into a training set (n = 147) and a test set (n = 76). The test set included 12 patients with abdominal aortic aneurysms; however, patients with existing vascular stents or bypass grafts were excluded to ensure the accuracy of the automated segmentation.³¹

The distribution of scanners is summarized below.

Siemens: Sensation 64 (test: 24, train: 51), Somatom Definition AS+ (test: 11, train: 31), and Somatom Definition Flash (train: 1).

GE: Optima CT660 (test: 24, train: 38) and LightSpeed 16 (test: 10, train: 16).

Toshiba: Aquilion ONE (test: 7, train: 8) and Aquilion PRIME (train: 2).

Additional relevant parameters are detailed in Table 1.

Table 1.

Parameters for the dataset.

Parameter	Test set (n = 76)	Training set (n = 147)
Age (in years)	70 (55–90)	72 (46–93)
Gender	48 M 28 F	97 M 50 F
Tube voltage (kV)	80–120	80–120
Slice thickness (mm)	1.9 (1.0–3.0)	2.1 (0.75–3.0)
Spacing (mm)	0.8 (0.63–1.12)	0.8 (0.62–1.10)

Abbreviations: M = male; F = female; kV = kilovolt; mm = millimeter.

The age, slice thickness, and spacing are reported as a mean with the range in brackets. Tube voltage is reported as a range.

Ground truth segmentations for the arterial system were generated through a hybrid pipeline of automated prediction and expert refinement. For the training set and 21 cases in the test set, initial arterial segmentations were produced using a previous neural network.³¹ For the remaining 55 test cases, the aorta was predicted by a separate network,³² while the branching arteries were manually delineated. The calcification segmentations were all done manually.

All segmentations underwent manual review and correction. The segmentation was performed by a vascular surgery resident with 4 years of post-medical school graduation training with guidance and consultation from a senior radiologist with over 20 years of experience.

Neural-network models

We employed the nnU-Net framework³³ for 3D semantic segmentation using the default plans with the 3D full-resolution configuration. Two ensemble models were trained, each comprising five independently trained models corresponding to a five-fold cross-validation scheme. In each fold, 117–118 of the 147 total volumes were used for training, and 29–30 for validation, ensuring that each volume was used as a validation sample exactly once. All models were trained for 1000 epochs per fold using nnU-Net’s default training parameters and preprocessing pipeline. One model was trained to segment arteries and calcifications separately, assigning label 1 to arteries and label 2 to calcifications. The second model was trained to segment the aorta, iliac arteries, and abdominal branches, with the aorta assigned label 1, iliac arteries assigned label 2, the mesenteric arteries label 3, and renal arteries assigned label 4. The models were trained using Tesla v100 16/32 GB GPUs in the University of Helsinki High Performance Computing platform Turso.

Shortening algorithm

To standardize the analysis across patients with varying anatomical lengths, we implemented an automated shortening algorithm as a post-processing step. After the neural network has generated a prediction for the arteries, it is then processed through an algorithm that shortens the side arteries to a determined length (2.5 cm) and truncates the aorta. It is similar to an earlier algorithm³¹ but with some improvements. The algorithm is detailed below, and the complete workflow for the method is illustrated in Figure 1.

1. Resample prediction segmentation into isotropic 1 mm spacing.

2. Run connected component analysis (CCA) on segmented side arteries.

3. For each component:

A. Define artery stump as an intersection of the component and twice dilated aortic segmentation.

B. Run CCA on stump, if there are components more than 2.5 cm apart from the largest components, label these components as background. Merge remaining components into a final stump segmentation.

C. Grow each stump segmentation by dilating the stump 25 times to reach desired length. Dilated voxels not intersecting with original artery segmentation of the stump are set as background.

D. Set label of the grown stump to match its corresponding artery segmentation.

4. Then it cuts off the aorta 1 cm above the highest side artery.

Figure 1.

Flowchart for the method. The CTA image input is processed by two separate nnU-Net segmentation models. One segments the different arteries and the other segments the calcifications. The artery segmentation is then processed by a shortening algorithm that shortens the side arteries to a determined length (2.5 cm). Combining the shortened artery segmentation and the calcification segmentation, different metrics of interest can be determined.

The algorithm was implemented as a Slicer module using Python 3.9.10 and Slicer 5.6.2. Additionally installed packages included scikit-image 0.24.0. A connectivity of 1 (6-connectivity in 3D) was used during CCA to define neighboring voxels.

Threshold-based comparison methods

To compare the neural network’s performance against traditional techniques, we utilized two threshold-based approaches.

First, an idealized method was used to determine the upper bound of thresholding performance. The idealized thresholding method uses ground truth segmentation to determine the optimal global threshold on a per-case basis. The method extracts the region covered by the ground truth arteries, tests a range of global thresholds at 1% intensity intervals, and selects the one that gives the highest Dice score when comparing the thresholded result to the ground truth. Because this requires prior ground-truth knowledge, it serves as a best-case reference rather than a viable clinical tool.

Second, a clinically realistic method was implemented by calculating the mean optimal threshold (470 HU in this case) across the entire training set and applying this fixed value to the test set.

Metrics for the neural network

Arterial segmentation performance was evaluated after the shortening algorithm was applied, using the Dice similarity coefficient, volumetric similarity, sensitivity, precision, and the Jaccard index.^34,35

Both predicted and ground truth volumes were processed through the standardized shortening algorithm, with the ground truths manually verified and refined post-processing. Performance was evaluated across four clinically relevant regions: the aorta, iliac, renal (LRA, RRA), and mesenteric arteries (SMA, CA). Calcification segmentation performance was assessed using identical metrics but was derived from the original predictions and ground truths.

Metrics for the calcification burden evaluation

We quantified calcification using three metrics: the total number of calcification clusters, the average cluster size, and the ratio of calcified volume to artery volume which is considered the main calcification burden score for this paper.

This patient-specific scoring framework was developed to account for the variability in arterial anatomy between patients and to mitigate the impact of fluctuations in neural network volume predictions. Because the calcification volume is anatomically constrained by the total volume of the artery, this score is naturally normalized to a range of 0 to 1.

The performance was evaluated against the ground truths for each artery group using the Pearson correlation, R² correlation, the Bland-Altman plot, and a regression plot. These statistical evaluations were used across all three metrics: the normalized burden score, cluster counts, and average cluster sizes.

Results

Performance of artery segmentation model

We first evaluated the artery segmentation model after the predictions had been processed by the shortening algorithm (Table 2). The model demonstrated robust performance across all arteries, with the highest scores observed in the aorta. Performance across the other arteries (iliac, renal, and mesenteric) was consistent and comparable to one another. Overall, the scores achieved by the model provide a good foundation for the subsequent calcification burden score analysis. Figure 2 provides a visual comparison example between the ground truth and the predicted arterial segmentations.

Table 2.

Segmentation metrics for the neural network models and the threshold-based methods.

Feature	Dice	VS	Sensitivity	Precision	Jaccard index
Abdominal aorta	0.953 ± 0.016 (0.906–0.981)	0.979 ± 0.016 (0.930–1.000)	0.959 ± 0.024 (0.867–0.991)	0.948 ± 0.033 (0.849–0.995)	0.910 ± 0.030 (0.828–0.963)
Iliac arteries	0.815 ± 0.101 (0.498–0.963)	0.933 ± 0.053 (0.711–1.000)	0.776 ± 0.112 (0.499–0.966)	0.866 ± 0.110 (0.456–0.991)	0.700 ± 0.137 (0.331–0.928)
Mesenteric arteries	0.815 ± 0.064 (0.602–0.940)	0.903 ± 0.074 (0.654–1.000)	0.883 ± 0.064 (0.667–0.984)	0.770 ± 0.117 (0.465–0.981)	0.692 ± 0.089 (0.431–0.887)
Renal arteries	0.784 ± 0.072 (0.557–0.913)	0.892 ± 0.081 (0.625–0.997)	0.841 ± 0.101 (0.411–0.989)	0.755 ± 0.119 (0.438–0.986)	0.651 ± 0.094 (0.386–0.840)
Calcification nnU-Net	0.772 ± 0.092 (0.512–0.903)	0.898 ± 0.086 (0.558–0.998)	0.771 ± 0.127 (0.360–0.958)	0.797 ± 0.123 (0.431–0.989)	0.636 ± 0.115 (0.344–0.823)
Calcification optimal thresholding	0.764 ± 0.117 (0.417–0.942)	0.887 ± 0.093 (0.655–0.997)	0.709 ± 0.162 (0.310–0.992)	0.848 ± 0.071 (0.612–0.978)	0.633 ± 0.149 (0.263–0.891)
Calcification global thresholding	0.560 ± 0.261 (0.007–0.928)	0.641 ± 0.290 (0.010–1.000)	0.676 ± 0.198 (0.213–0.985)	0.664 ± 0.357 (0.004–1.000)	0.431 ± 0.241 (0.004–0.866)

Abbreviations: VS = volumetric similarity.

The metrics are of the form mean ± standard deviation (range).

The arteries were evaluated after being processed by the shortening algorithm.

Figure 2.

Comparison between the ground truth artery segmentation (left) and the predicted artery segmentation (right). The shortening algorithm has been used to trim the iliac, mesenteric, and renal arteries to approximately 2.5 cm.

Calcification segmentation: neural network versus thresholding

The comprehensive evaluation of the calcification can be seen in Table 2 (segmentation metrics), Table 3 (correlations), and Figures 4–7 (Bland-Altman plots and regression lines). It is important to note that while the pure segmentation metrics compare predictions directly against the calcification ground truth, the calcification burden evaluation metrics are composite measures; they are influenced by both the calcification segmentation and the accuracy of the underlying shortened arterial segmentation.

Table 3.

Correlations between the scores (calcification burden scores, calcification cluster sizes, and counts) determined using ground truth segmentations and the scores determined using the neural network arterial predictions and the three methods for segmenting calcifications.

Feature	nnU-Net (PCC)	nnU-Net (R²)	Optimal thresholding (PCC)	Optimal thresholding (R²)	Clinical thresholding (PCC)	Clinical thresholding (R²)
Aorta calcification burden score	0.955	0.907	0.958	0.885	0.162	−19.8
Iliac arteries calcification burden score	0.943	0.888	0.932	0.855	0.443	−0.987
Mesenteric arteries calcification burden score	0.869	0.654	0.891	0.705	0.216	−9.13
Renal arteries calcification burden score	0.933	0.773	0.930	0.806	0.417	−5.45
All arteries combined calcification burden score	0.954	0.905	0.958	0.887	0.187	−13.3
Aorta calcification cluster count	0.781	0.445	0.607	−8.13	0.118	−53.0
Iliac arteries calcification cluster count	0.802	0.590	0.779	0.353	0.302	−14.7
Mesenteric arteries calcification cluster count	0.691	0.464	0.508	−4.05	0.107	−28.3
Renal arteries calcification cluster count	0.800	0.636	0.771	0.353	0.191	−9.09
Aorta average calcification cluster size	0.526	−7.76	0.59	−0.037	0.059	−2094
Iliac arteries average calcification cluster size	0.445	−0.16	0.48	0.123	0.165	−2.35
Mesenteric arteries average calcification cluster size	0.926	0.673	0.654	0.422	0.187	−4.02
Renal arteries average calcification cluster size	0.803	0.024	0.776	−0.118	0.466	−16.8

Abbreviations: PCC = Pearson correlation coefficient.

The scores here are composite scores utilizing both arterial and calcification segmentations.

Figure 3.

Comparison between the ground truth calcification segmentation (left) and the predicted segmentation by different methods (neural network prediction middle left, optimized threshold middle right, and the clinical threshold on the right). This illustrates the limitation of the clinical method as the HU values of the arteries were higher than 470 HU.

Figure 4.

Bland-Altman plots for the datapoints. It shows the plots for the scores of different arteries; aorta, iliac arteries, mesenteric arteries, and renal arteries and then the combined score of the arteries between the ground truth scores and both the optimal thresholding method score and the neural network method score. The score is the volume of calcification divided by the volume of the artery. The dotted middle line shows the mean difference between the ground truth score and the predicted score, and the outer dotted lines show the mean difference ±1.96*SD (SD = standard deviation of the difference).

Figure 5.

The figure shows linear regression plots for the calcification scores of different arteries; aorta, iliac arteries, mesenteric arteries, and renal arteries and then the combined score of the arteries between the ground truth and the predicted score values for both the optimal thresholding method score and the neural network method score. The score is the volume of calcification divided by the volume of the artery. The colored area represents 95% confidence intervals, and the dotted line represents the line of unity. In Seaborn, the confidence interval is estimated using a bootstrap.

Figure 6.

Bland-Altman plots for the calcification clusters. It shows the plots for the amount and sizes of calcification clusters of different arteries; aorta, iliac arteries, mesenteric arteries, and renal arteries between the ground truth clusters and both the optimal thresholding method clusters and the neural network method clusters. The average size of the clusters is in mm^3. The dotted middle line shows the mean difference between the ground truth score and the predicted score, and the outer dotted lines show the mean difference ±1.96*SD (SD = standard deviation of the difference).

Figure 7.

The figure shows linear regression plots for the amount and sizes of calcification clusters of different arteries; aorta, iliac arteries, mesenteric arteries, and renal arteries between the ground truth clusters and both the optimal thresholding method clusters and the neural network method clusters. The average size of the clusters is in mm^3. The colored area represents 95% confidence intervals, and the dotted line represents the line of unity. In Seaborn, the confidence interval is estimated using a bootstrap.

In terms of segmentation metrics (Table 2), the neural network method slightly outperformed the optimized thresholding method in most categories. The notable exception was precision, where the optimized thresholding method was superior, whereas the neural network demonstrated greater sensitivity. The neural network seems to offer a more balanced compromise between sensitivity and precision. In contrast, the clinically realistic thresholding method failed to reach the performance levels of either the neural network or the optimized approach (Tables 2 and 3). A visual comparison between the calcification segmentations can be seen in Figure 3.

Next is the analysis of the accuracy of the calcification burden score. Bland-Altman analysis (Figure 4) revealed that both the neural network and optimized thresholding methods tended to underestimate calcification scores in the aorta and iliac arteries, though the bias was more pronounced in the optimized thresholding method. Conversely, in the smaller mesenteric and renal arteries, the bias shifted toward overestimation. Despite these trends, most data points for both methods fell well within the confidence intervals.

Regression analysis (Figure 5) showed that the neural network’s predictions for the aorta and iliac arteries aligned closely with the unity line. Although the correlation was slightly lower in the smaller side arteries (mesenteric and renal), the combined artery volume score had high correlation as it is dominated by the larger aortic and iliac volumes.

While the optimized thresholding method generally showed slightly higher Pearson correlation coefficients (PCCs) in Table 3 than the neural network (except in the iliac arteries), a paired t-test confirmed no statistically significant difference between the two methods regarding either Dice scores or final calcification scores (p > .05). Overall, the high PCC for the calcification burden score across all arteries supports the feasibility of this method for the evaluation of the calcification burden.

Finally, we analyzed the morphology of the calcifications (Figures 6 and 7). The models exhibited distinct behaviors: the optimized thresholding tended to overestimate the amount of calcification clusters. Conversely, the neural network tended to underestimate the count while overestimating the average size of clusters. This suggests that the network merges adjacent small calcifications into single, larger clusters. In Table 3, we see the correlation between the prediction and the ground truths favors the neural network model in the cluster counts and generally the optimized thresholding method in cluster sizes.

The correlations for cluster counts and sizes were significantly weaker than those for the total calcification score (Table 3). This finding supports the use of the volume-normalized burden score as a more robust and reproducible metric for the evaluation of the calcification burden.

Discussion

This study compared a neural network-based method for quantifying aortic, iliac, renal, and mesenteric artery calcification with an idealized threshold-based approach and a more clinically achievable automated thresholding method. Our results demonstrate that the neural network and idealized thresholding achieved similar outcomes, with high correlations to the ground truth, while the automated clinical thresholding performed significantly worse. This suggests that a deep-learning approach can match or even slightly improve upon idealized thresholding, indicating the feasibility of automated calcification scoring in CTA. Consequently, this method offers a potential alternative to traditional non-contrast Agatston scoring,²⁵ potentially reducing patient radiation exposure by eliminating the need for additional unenhanced scans.

We utilized separate neural networks to segment calcifications and arteries to better determine which artery the calcification belongs to. In the case where the model would have predicted both it would be more difficult to make that determination, especially on the borders of the different arteries. We opted for a post-processing shortening algorithm over training on pre-shortened data mostly due to its inherent flexibility; for example, the length of side arteries can be adjusted from 2.5 cm to 3.5 cm without requiring model retraining. Furthermore, to account for inter-patient anatomical variability and variations in artery segmentation quality, we implemented a volume-normalized scoring framework that enables more consistent comparison of calcification burden across different artery types and patients. While both artery and calcification segmentations influence the final score, normalization by artery volume helps reduce the bias caused by the artery segmentation.

In addition to volumetric quantification of the calcification, we also evaluated the number and average size of calcification clusters within each artery group. This analysis provides a complementary perspective on the morphology of vascular calcification, which may have different clinical relevance compared to pure calcification volume. However, the correlations suggest that the current method is not optimal for consistently and accurately evaluating cluster sizes and counts.

The failure of the simple clinical thresholding algorithm further illustrates the need for case-specific methods. We tested more nuanced approaches, such as Otsu-thresholding and highest-HU percentiles, but neither improved Dice scores. A primary reason for the poor correlation in clinical thresholding was that in cases where the entirety of the artery had higher than average HU values, the algorithm occasionally misidentified the entire artery as calcification (Figure 3).

A key element of this study is the division of the abdominal arterial system into four anatomically and clinically relevant groups: the aorta, iliac arteries, renal arteries, and mesenteric arteries (coeliac and superior mesenteric arteries). This separation enables more detailed, region-specific analysis of calcification patterns and reflects the varying clinical implications across vascular territories. By consistently labeling and standardizing the segment lengths of these arteries, our method supports the development of more precise inter-patient comparisons and artery-specific or composite risk scores tailored to clinical context.

Despite these findings, some limitations of this study must be noted. Both the training and the test data were obtained from a single center, leaving the model’s generalizability to other patient populations unclear. Vascular stents and bypass grafts, which can interfere with calcification and artery segmentation,³¹ were excluded from the test set. However, future improvements in imaging, such as the photon-counting detectors, may eventually allow for easier differentiation between stents and calcifications.³⁶ Moreover, the clinical relevance of the calcification burden score requires further validation before it can be applied in clinical practice. While the method currently provides four artery-specific scores, the next possible step would be to create a composite calcification score potentially with weighted contributions based on the clinical significance of each territory. Further research is necessary to validate the correlation and predictive value of the artery-specific scores regarding clinical outcomes. These scores hold significant potential as predictive markers for complications such as renal insufficiency, mesenteric ischemia, and postoperative bowel resection complications, warranting further investigations. The predictive value of different factors has been studied previously using machine learning methods.³⁷

In conclusion, our results demonstrate that a neural network-based system offers a robust, observer-independent alternative to threshold-based techniques for calcification assessment in CTA. With further validation and clinical integration, this method may enhance cardiovascular risk assessment and support treatment planning in patients with peripheral artery disease.

Footnotes

Acknowledgments

The authors wish to thank the Finnish Computing Competence Infrastructure (FCCI) for supporting this project with computational and data storage resources.

ORCID iDs

Johannes Halkoaho

Oskari Niiranen

Ethical considerations

Author contributions

J.H. contributed to the design of the work, the analysis and interpretation of data, the creation of new software used in the work, and drafted the work. J.H. was the main writer of the study, performed data analysis, contributed to data acquisition, and contributed to algorithm development and study design. O.N. contributed to the acquisition of data and performed ground truth segmentation. T.K. contributed to the creation of new software used in the work. T.K. was the main algorithm developer and contributed to writing. A.R. contributed to the creation of new software used in the work and the acquisition of data. A.R. trained the neural networks and provided the predictions and contributed to writing. E.S. and S.S. contributed to the conception and design of the work and substantially revised it. E.S. and S.S. supervised the project, provided critical input on study design, and revised the manuscript. H.H. and M.K. contributed to the conception and design of the work and substantially revised it. H.H. and M.K. supervised and conceived the project, provided critical input on study design and interpretation, and revised the manuscript. H.H. and M.K. were responsible for resources and project administration. All authors made substantial contributions to the conception and design of the work. All authors critically revised the manuscript for important intellectual content and approved the final version of the manuscript.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study received funding from Helsinki University Hospital (MK: TYH2024228, M780025014) along with Federal Grant Satasairaala and Finnish Culture Foundation Satakunta fund (grant numbers: 75212239 and 7522150).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The datasets analyzed during the current study are not publicly available due to Finland’s regulations, but some parts of the analysis are available from the corresponding author on reasonable request*.

References

Fowkes

FGR

Rudan

, et al. Comparison of global estimates of prevalence and risk factors for peripheral artery disease in 2000 and 2010: a systematic review and analysis. Lancet 2013; 382: 1329–1340. https://doi.org/10.1016/S0140-6736(13)61249-0

Newman

Sutton-Tyrrell

Vogt

, et al. Morbidity and mortality in hypertensive adults with a low ankle/arm blood pressure index. JAMA 1993; 270: 487–489.

Ankle Brachial Index Collaboration Fowkes

FGR

Murray

, et al. Ankle brachial index combined with Framingham Risk Score to predict cardiovascular events and mortality: a meta-analysis. JAMA 2008; 300: 197–208. https://doi.org/10.1001/jama.300.2.197

Mendis

Lindholm

Anderson

, et al. Total cardiovascular risk approach to improve efficiency of cardiovascular prevention in resource constrain settings. J Clin Epidemiol 2011; 64: 1451–1462. https://doi.org/10.1016/j.jclinepi.2011.02.001

Lin

Olson

Johnson

, et al. The ankle-brachial index for peripheral artery disease screening and cardiovascular disease prediction among asymptomatic adults: a systematic evidence review for the U.S. Preventive Services Task Force. Ann Intern Med 2013; 159: 333–341. https://doi.org/10.7326/0003-4819-159-5-201309030-00007

Fowkes

FGR

Murray

Butcher

, et al. Development and validation of an ankle brachial index risk model for the prediction of cardiovascular events. Eur J Prev Cardiol 2014; 21: 310–320. https://doi.org/10.1177/2047487313516564

Saeedi

Petersohn

Salpea

, et al. Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: results from the International diabetes federation diabetes Atlas, 9th edition. Diabetes Res Clin Pract 2019; 157: 107843. https://doi.org/10.1016/j.diabres.2019.107843

Roth

Mensah

Johnson

, et al. Global burden of cardiovascular diseases and risk factors, 1990-2019: update from the GBD 2019 Study. J Am Coll Cardiol 2020; 76: 2982–3021. https://doi.org/10.1016/j.jacc.2020.11.010

Zhang

Sun

Wang

, et al. A nomogram to predict the risk of colorectal anastomotic leakage combining inflammatory-nutritional and abdominal aorta calcium index. Front Surg 2022; 9: 1008448. https://doi.org/10.3389/fsurg.2022.1008448

10.

Hoek

Edomskis

Menon

, et al. Arterial calcification is a risk factor for anastomotic leakage after esophagectomy: a systematic review and meta-analysis. Eur J Surg Oncol 2020; 46: 1975–1988. https://doi.org/10.1016/j.ejso.2020.06.019

11.

Aoyama

Fukumoto

Shigita

, et al. Arteriosclerosis is a major predictor of small bowel vascular lesions. Dig Dis Sci 2018; 63: 723–730. https://doi.org/10.1007/s10620-018-4930-x

12.

Sadat

Mariam

NBG

Usman

, et al. Association between abdominal visceral artery calcification and all-cause mortality-a computerized tomography imaging-based longitudinal follow-up study. Angiology 2019; 70: 237–243. https://doi.org/10.1177/0003319718796889

13.

Koyanagi

Ozawa

Ninomiya

, et al. Association between indocyanine green fluorescence blood flow speed in the gastric conduit wall and superior mesenteric artery calcification: predictive significance for anastomotic leakage after esophagectomy. Esophagus 2021; 18: 248–257. https://doi.org/10.1007/s10388-020-00797-8

14.

Lee

S-L

Jeon

C-H

Park

K-B

, et al. Association between vascular calcification and esophagojejunal anastomotic complications after total gastrectomy for gastric cancer: a propensity-matched Study. Curr Oncol 2022; 29: 3224–3231. https://doi.org/10.3390/curroncol29050262

15.

Chang

D-H

Brinkmann

Smith

, et al. Calcification score versus arterial stenosis grading: comparison of two CT-based methods for risk assessment of anastomotic leakage after esophagectomy and gastric pull-up. Ther Clin Risk Manag 2018; 14: 721–727. https://doi.org/10.2147/TCRM.S157352

16.

De’Ath

Oakland

Brohi

. CT screened arterial calcification as a risk factor for mortality after trauma. Scand J Trauma Resusc Emerg Med 2016; 24: 120. https://doi.org/10.1186/s13049-016-0317-1

17.

Postaire

Abet

Montigny

, et al. Does the degree of calcification of the celiac trunk and superior mesenteric artery on preoperative computerized tomography predict the risk of anastomotic leak after right colectomy? A single center retrospective study. J Visc Surg 2019; 156: 191–195. https://doi.org/10.1016/j.jviscsurg.2018.10.006

18.

Idil Soylu

Avcıoglu

Uzunkaya

, et al. Evaluation of mesenteric artery disease in patients with severe aortic valve stenosis. J Investig Med 2021; 69: 719–723. https://doi.org/10.1136/jim-2020-001549

19.

Tong

Xie

Song

, et al. Is abdominal vascular calcification score valuable in predicting the occurrence of colorectal anastomotic leakage? A meta-analysis. Int J Colorectal Dis 2020; 35: 641–653. https://doi.org/10.1007/s00384-020-03513-1

20.

Takeyoshi

Shirasaka

Shibagaki

, et al. Mesenteric ischemia after cardiac surgery in dialysis patients: an overlooked risk factor. Heart Surg Forum 2022; 25: E732–E738. https://doi.org/10.1532/hsf.4859

21.

Miura

Kurimoto

Iba

, et al. Quantitative evaluation of superior mesenteric artery calcification in hemodialysis patients undergoing aortic valve replacement. Gen Thorac Cardiovasc Surg 2020; 68: 1252–1259. https://doi.org/10.1007/s11748-020-01352-z

22.

Koo

Allison

Criqui

, et al. The association between liver fat and systemic calcified atherosclerosis. J Vasc Surg 2020; 71: 204–211.e4. https://doi.org/10.1016/j.jvs.2019.03.044

23.

Iijima

Kotani

Sakuma

, et al. Risk factors for celiac artery and superior mesenteric artery stenosis in preoperative lumbar and thoracolumbar spinal surgery patients. J Orthop Sci 2023; 28: 1214–1220. https://doi.org/10.1016/j.jos.2022.10.004

24.

Lareyre

Augène

Massalou

, et al. Vascular calcifications are associated with increased mortality in patients with acute mesenteric ischemia. Ann Vasc Surg 2021; 72: 88–97. https://doi.org/10.1016/j.avsg.2020.08.086

25.

Graffy

Liu

O’Connor

, et al. Automated segmentation and quantification of aortic calcification at abdominal CT: application of a deep learning-based algorithm to a longitudinal screening cohort. Abdom Radiol 2019; 44: 2921–2928. https://doi.org/10.1007/s00261-019-02014-2

26.

Isgum

van Ginneken

Olree

. Automatic detection of calcifications in the aorta from CT scans of the abdomen. 3D computer-aided diagnosis. Acad Radiol 2004; 11: 247–257. https://doi.org/10.1016/s1076-6332(03)00673-1

27.

Park

Kwon

S-S

Song

, et al. Deep learning based automatic quantification of aortic valve calcification on contrast enhanced coronary CT angiography. Sci Rep 2025; 15: 8472. https://doi.org/10.1038/s41598-025-93744-5

28.

Fornasari

Kuntz

Martini

, et al. Objective methods to assess aorto-iliac calcifications: a systematic review. Diagnostics 2024; 14: 1053. https://doi.org/10.3390/diagnostics14101053

29.

Buijs

RVC

Leemans

Greuter

, et al. Quantification of abdominal aortic calcification: inherent measurement errors in current computed tomography imaging. PLoS One 2018; 13: e0193419. https://doi.org/10.1371/journal.pone.0193419

30.

Holcombe

Horbal

Ross

, et al. Variation in aorta attenuation in contrast-enhanced CT and its implications for calcification thresholds. PLoS One 2022; 17: e0277111. https://doi.org/10.1371/journal.pone.0277111

31.

Halkoaho

Niiranen

Salli

, et al. Quantifying the calcification of abdominal aorta and major side branches with deep learning. Clin Radiol 2024; 79: e665–e674. https://doi.org/10.1016/j.crad.2024.01.023

32.

Kesävuori

Kaseva

Salli

, et al. Deep learning-aided extraction of outer aortic surface from CT angiography scans of patients with Stanford type B aortic dissection. Eur Radiol Exp 2023; 7: 35. https://doi.org/10.1186/s41747-023-00342-z

33.

Saa

et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021; 18: 203–211. https://doi.org/10.1038/s41592-020-01008-z

34.

Taha

Hanbury

. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC Med Imag 2015; 15: 29. https://doi.org/10.1186/s12880-015-0068-x

35.

Bellal

El Fkihi

Cengiz

, et al. A comprehensive survey on deep learning in abdominal imaging: datasets, techniques, and performance metrics. IEEE Access 2025; 13: 79894–79914. https://doi.org/10.1109/access.2025.3566627

36.

De Beukelaer

Wuyts

, et al. Photon-counting detector CTA to assess intracranial stents and flow diverters: an in vivo study with ultrahigh-resolution spectral reconstructions. European Radiology Experimental 2025; 9: 10. https://doi.org/10.1186/s41747-025-00550-9

37.

Song

Liu

Wang

, et al. Interpretable machine learning prediction model for major adverse cardiovascular events in patients with peripheral artery disease. J Vasc Surg 2025; 1: S0741521425010948.

Neural network assessment of aortic,iliac,renal,and mesenteric artery calcification in CTA: Normalized scoring framework and comparison to threshold-based method

Abstract

Background

Purpose

Material and Methods

Results

Conclusion

Keywords

Introduction

Materials and methods

Data

Neural-network models

Shortening algorithm

Threshold-based comparison methods

Metrics for the neural network

Metrics for the calcification burden evaluation

Results

Performance of artery segmentation model

Calcification segmentation: neural network versus thresholding

Discussion

Footnotes

Acknowledgments

ORCID iDs

Ethical considerations

Author contributions

Funding

Declaration of conflicting interests

Data Availability Statement

References