Cardiovascular diseases are the leading cause of death
among patients with chronic kidney disease (CKD) and those
undergoing hemodialysis. In these patients, acute myocardial
infarction, fatal arrhythmias, and heart failure are among the
primary causes of cardiovascular-related mortality.
Abdominal aortic calcification (AAC) is an important
morphological marker of vascular pathology, reflecting the
extent of vascular calcification, and is considered a significant
predictor of cardiovascular events. However, the assessment
of AAC scores relies on the experience and subjective
judgment of physicians, leading to variability in results due to
differing standards among doctors, causing confusion and
requiring significant manpower. In light of this, the current
paper aims to utilize artificial intelligence models to precisely
measure the degree of abdominal aortic calcification in
abdominal X-rays, and use this score to predict the risk of
death, cardiovascular death, major cardiovascular obstruction,
and peripheral vascular obstruction events in hemodialysis
patients.
The main goal of this paper is to develop an artificial
intelligence-based model for the accurate assessment of
abdominal aortic calcification in abdominal X-rays.
Additionally, this paper will explore the association between
AAC scores and the risk of central vascular events and death
in hemodialysis patients, aiming to establish a combined
model of related risk factors and prognostic prediction models
for clinical application. Developing an AAC network model
that can calculate the abdominal aortic calcification score
more accurately than traditional medical imaging reading has
significant implications for improving the quality and clinical
value of abdominal X-ray reports. By enabling early and
precise assessment of abdominal aortic calcification,
physicians can intervene earlier in treatment, thereby reducing
the mortality rate from chronic hemodialysis, cardiovascular
death, and the occurrence of major adverse cardiovascular
events and peripheral vascular events. Furthermore, as a
powerful predictive indicator, the abdominal aortic
calcification score can help physicians better assess patient
risk and devise personalized treatment plans.
Vascular calcification (VC) is a common phenomenon of
vascular aging that poses a significant threat to cardiovascular
health. The abdominal aorta, one of the body's principal
arteries, is responsible for transporting blood from the heart to
the lower body. The presence of Abdominal Aortic
Calcification (AAC) not only restricts blood flow but also
reduces arterial elasticity, thereby increasing the risk of
cardiovascular diseases, cerebrovascular diseases, and
peripheral vascular diseases, potentially leading to death.
Patients with chronic kidney disease (CKD) are particularly
prone to VC, especially AAC, with a prevalence rate as high
as 60% [1]. The high incidence of VC among CKD patients is
partly due to traditional risk factors such as aging,
hypertension, type 2 diabetes, dyslipidemia, and smoking, as
well as CKD-related non-traditional risk factors, including
uremia, chronic inflammation, vascular aging, and
hyperphosphatemia [2].
The calcification of the abdominal aorta affects a wide
range, from small to medium and large arteries. Assessing the
AAC score as determined by physicians serves as an
important indicator for evaluating vascular calcification and is
crucial for assessing the overall vascular health. Therefore, an
objective assessment of the AAC score is a significant task for
clinical physicians, aiding in the early detection and
formulation of treatment plans to reduce the risk of related
diseases. The method of calculating the calcification score is
illustrated in Fig. 2, where the abdominal aorta is divided into
Calculation of Aortic Arch Calcification Degree in Hemodialysis
Patients Using Deep Learning
CHUNG-KUAN WU1, CHE-YU CHIANG2, JUN-WEI HSIEH3,*
1Nephrology Shin Kong Wu Huo-Shih Memorial Hospital Taipei, TAIWAN
2Department of Computer Science and Information Engineering National Taiwan Ocean University Keelung, TAIWAN
3Institute of Intelligent Computation National Yang-Ming Chiao-Tung University Tainan, TAIWAN
Abstract: - Abdominal Aortic Calcification (AAC) is a common form of vascular calcification closely associated with
atherosclerosis and serves as an important marker for measuring increased risk of cardiovascular, cerebrovascular, and
peripheral vascular diseases. Particularly in patients with Chronic Kidney Disease (CKD) and those undergoing dialysis,
the risk of AAC significantly increases due to a combination of traditional and non-traditional risk factors. Therefore,
developing a rapid and accurate method to assess the extent of AAC is crucial for preventing the progression of vascular
calcification and the associated risk of cardiovascular diseases. Dialysis patients are required to undergo an abdominal X-
ray annually, and the degree of calcification of the abdominal aorta is assessed manually through these X-ray images.
However, these methods have limitations in identifying subtle calcifications in the abdominal aorta and the assessment
process is time-consuming and depends on the experience and subjective judgment of physicians. To overcome these
limitations, we propose a new method that incorporates deep learning technology to improve the accuracy of assessing the
extent of AAC. Our method utilizes CNN models and attention modules to enhance the model's ability to recognize features
of abdominal aortic calcification.
Key-words: - Abdominal Aortic Calcification, Residual Network, Deep Learning, CKD, Cross-scale Attention.
Received: December 21, 2023. Revised: August 14, 2024. Accepted: September 19, 2024. Published: October 8, 2024.
* Corresponding author.
1. Introduction
2. Related Works
International Journal of Applied Sciences & Development
DOI: 10.37394/232029.2024.3.17
Chung-Kuan Wu, Che-Yu Chiang, Jun-Wei Hsieh
E-ISSN: 2945-0454
171
Volume 3, 2024
Fig. 1. The flow chart of AAC (Abdominal Aortic Calcification) Network
four sections (L1-L4) based on the position of four vertebrae,
with each section further divided into the Posterior Wall(PW)
and the Anterior Wall (AW), totaling eight regions.
The assessment of AAC primarily utilizes two semi-
quantitative scoring systems: the Kaupplia AAC-24 point and
AAC-8 point scoring scales [3,4]. The AAC-24 point scoring
system rates the degree of calcification coverage on the
anterior and posterior walls of the abdominal aorta across four
segments in front of the L1 to L4 lumbar vertebrae, using a
scoring standard of 0 to 3, with a total score ranging from 0 to
24. The AAC-8 point scoring scale, a simplified version,
scores by quantifying the total length of calcification on the
anterior and posterior walls in front of the L1 to L4 lumbar
vertebrae, with total scores ranging from 0 to 8. Due to the
convenience of AAC-8 and its sufficient clinical applicability,
this paper primarily adopts AAC-8 as the accuracy
classification standard, while AAC-24 is used for comparison
with other research.
Fig. 2. Schema for AAC-24 point scale scores.(a) 0 (b) 2 (c) 6 (d) 15 .
Schema for AAC-8 point scale scores.(a) 0 (b) 2 (c) 3 (d) 8 .
In recent years, with the rapid development and success of
Convolutional Neural Networks (CNNs), many researchers
have applied deep learning techniques to medical-related
topics. However, the features of most medical samples are not
as rich and diverse as objects in natural scenes. For example,
vascular calcification in X-ray images looks very similar to
normal bone, and blood vessels without calcification are
completely invisible. As a result, manual inspection or
traditional vision techniques may lead to misclassification.
This paper proposes a new CNN architecture for automatically
calculating the degree of abdominal aortic calcification in
patients undergoing hemodialysis
To address the issues above, this paper introduces a
network architecture called AAC-Net (Abdominal Aortic
Calcification Network), which combines attention
mechanisms with a CNN model. By utilizing deep learning
technology for more detailed analysis of image data, this
method can help physicians more accurately identify and
assess the degree of abdominal aortic calcification, even at an
early stage. This could contribute to early intervention,
reducing the risk of cardiovascular diseases caused by
vascular calcification, thereby improving patient prognosis
and quality of life.
The flow chart of AAC-Net is depicted in Fig. 1. This
AAC network can directly analyze and detect the position of
the spine from spinal X-ray images, then locate the position of
the abdominal aorta. It enhances features through feature
extraction and various attention modules, followed by an MLP
(Multi-Layer Perceptron) network that classifies the degree of
calcification in patients. The AAC network architecture is
divided into detection and classification steps. Initially, our
previously proposed object detector, PRB-FPN [5], is used for
detecting the position of the spine. It can detect very small
objects more effectively than YOLO v4 [6], through which the
position of the abdominal aorta is determined. Fig. 3 shows
the results of abdominal aorta detection using PRB-Net. In the
classification step, a model with ResNet [7] as the backbone
is used to extract features. To enhance the discriminative
capability of features, we propose COP-Net (Cross-scale
Overlapping Patch-based Network) to use features as an
additional attention, inputting the weight map produced by
COP-Net into the model. The weight map strengthens the
calcification areas focused on by the classification network.
Finally, classification and calculation of the calcification score
are performed using a three-layer FC with ReLU. Our AAC
network architecture also adopts the AAC-8 for calculating
the abdominal aortic calcification score, achieving a top
accuracy rate of 79.24%.
Fig. 3. Abdominal aorta detection result of PRB-FPN
3. Method
International Journal of Applied Sciences & Development
DOI: 10.37394/232029.2024.3.17
Chung-Kuan Wu, Che-Yu Chiang, Jun-Wei Hsieh
E-ISSN: 2945-0454
172
Volume 3, 2024
Fig. 4. Architecture of COP-Net (Cross-scale Overlapping Patch-based Network).
Fig. 4 illustrates the architecture of COP-Net, starting with
the model using ResNet [7] as the backbone to obtain features.
Then, through the SCA (Scale-aware Channel Attention)
module, the feature maps of different channels are weighted.
Next, using the PCA (Patch-based Cross-scale Attention)
Module, the feature maps are divided into different patches.
Through different scales of feature maps and self-attention,
these feature maps undergo Query, Key, Value weighting to
obtain more discriminative feature maps. Finally, AAC
detection is performed using a 1x1 convolution mechanism,
detecting the calcified pixel positions and generating a
heatmap of the abdominal aorta AAC. The classification of
abdominal aortic X-ray images and calculation of the
calcification score are done using a three-layer Fully-
Connected Network (or Multiple-Layer Perceptron), with the
images outputting three categories representing no
calcification, calcification on one side of the vessels, and
calcification on both sides of the vessels.
We adopted ResNet18 [7] as the feature extraction module
and experimented with various attention modules, including
CBAM [8], SimAM [9], and Shuffle Attention(SA) [10],
among others. These attention modules significantly enhance
the discriminative ability of the backbone network for AAC
classification. Fig. 5 shows our method for AAC classification,
dividing a segment of the patient's spine corresponding to
vascular X-ray images into three types of grading results. The
results are shown in Figure 6. Table I demonstrates the
performance of these three attention modules. Upon
examination, we found that although both SimAM [9] and
CBAM [8] aim to enhance the model's feature discriminative
power, their performance in this task was slightly inferior to
that of ResNet18 [7]. This result could be attributed to several
factors. First, the SimAM [9]and CBAM [8] attention module
may cause the model to overly focus on channel features while
neglecting spatial information, which is crucial for the
detection of vascular calcification. The introduction of the
Shuffle Attention (SA) [10] module, which combines spatial
and channel attention mechanisms, enhances feature
representation by rearranging feature channels and spatial
attention to facilitate cross-channel information exchange.
When calculating the calcification score for AAC, there
can be individual calcifications on the posterior wall (PW) and
anterior wall (AW). Without separating the inputs into PW
and AW, the classifier can become confused. To further
increase the accuracy of the previous classification model, we
divided the input into three parts: FS (Full-Size) for the entire
image, PW (Posterior Wall) portion, and AW (Anterior wall)
portion. The model outputs four categories: no calcification,
AW calcification, PW calcification, and calcification on both
sides of the vessels. This approach has two advantages: 1.
More discriminative features: Since the scoring items are
composed of AW and PW, inputting them separately provides
the model with additional information. 2. Finer classification:
The classification task becomes four categories, which allows
for the assessment of calcification in AW and PW individually
(secondary tasks), thereby enhancing the accuracy of the
model's classifications. Fig. 6 is the flowchart of this method,
and Table II shows the accuracy of calcification score
calculations for this method. Compared to Table I, there is a
significant improvement in accuracy, and the highest accuracy
is achieved when the SimAM [9] attention mechanism is
employed.
Fig. 5. Single abdominal aortic section images as input features.
Fig. 6. Classification results of AAC score for aortic area corresponding to
a single spinal segment.
Indeed, we can design a network architecture to detect the
situation of abdominal aortic calcification and then use this
calcification condition as a basis for feature weighting. We use
a COP-Net (Cross-scale Overlapping Patch-based Network),
which outputs the condition of aortic calcification into a
feature heatmap. The core of COP-Net lies in its two
4. Experiments and Results
International Journal of Applied Sciences & Development
DOI: 10.37394/232029.2024.3.17
Chung-Kuan Wu, Che-Yu Chiang, Jun-Wei Hsieh
E-ISSN: 2945-0454
173
Volume 3, 2024
innovative modules: the Scale-aware Channel Attention (SCA)
module and the Patch-based Cross-scale Attention (PCA)
module. The introduction of the SCA module enables our
method to assess the importance of channels at different scales.
The PCA Module slices the feature map into different patches,
and then, using feature maps of different scales, employs self-
attention to weight these feature maps with Query, Key, Value,
obtaining more discriminative feature maps. Finally, a 1×1
convolution mechanism is used to generate a channel attention
feature map that is more global and comprehensive. The
heatmaps generated by COP-Net are used as weights, further
input into our classification model. Using these heatmaps as
weights for training the classification model means we can
guide the classification model to focus on those areas that are
most decisive for the final diagnosis. Fig. 8 shows the process
of converting an abdominal aortic X-ray image into an arterial
calcification heatmap using COP-Net; (c) shows the original
image multiplied by the heatmap to obtain an enhanced feature
map, and then features are extracted with ResNet and
calcification scores are calculated with an MLP. The entire
process is depicted in Fig. 9.
TABLE I. ACCURACY OF AAC CALCULATION WITH RESNET AND
VARIOUS ATTENTION MODULES
Acc(%)
66.67
64.67
65.33
69.00
Fig. 7. Segmenting a single abdominal aortic section into three aarts as
input features: full-size Image, posterior wall (PW), and anterior wall
(AW).
TABLE II. ANALYSIS OF AAC SCORE CALCULATION ACCURACY
USING SINGLE ABDOMINAL AORTIC SECTION SEGMENTED INTO FULL-SIZE
IMAGE, ANTERIOR WALL, AND POSTERIOR WALL AS INPUT FEATURES.
Model
Acc(%)
ResNet18
71.50
ResNet34
65.00
ResNet18+SimAM
75.75
ResNet18+CBAM
74.75
ResNet18+SA
72.50
Table III presents an accuracy analysis of AAC score
calculation using arterial calcification heatmaps as feature
weighting. The second column represents the core idea of
multiplying the original image by the heatmap generated by
COP-Net before input, utilizing the heatmap to highlight key
feature areas. These areas' information is directly integrated
into the original image in an attentional manner, thereby
enhancing the model's accuracy in calculating AAC scores.
By multiplying the heatmap with the original image, we
effectively focus the model's attention on key abnormal areas
such as abdominal aortic calcification. This method not only
preserves the detailed information in the original image but
also enhances the model's ability to recognize markers of
specific health conditions. The third column represents the
accuracy of AAC score calculation by the network when the
original image and the heatmap generated by COP-Net are
concatenated. The results show that concatenation can more
precisely identify the presence and extent of abdominal aortic
calcification, thus achieving higher classification accuracy.
Fig. 8. Transforming (a) into a heatmap of arterial calcification using COP-
Net as shown in (b); (c) the original image multiplied by the heatmap
to obtain an enhanced feature map.
Fig. 9. Using COP-Net to transform abdominal aortic images into heatmaps
of arterial calcification, followed by multiplying the original images by the
heatmap to obtain enhanced feature maps. Then, features are extracted using
ResNet and calcification scores are calculated with MLP.
TABLE III. ANALYSIS OF ACCURACY IN AAC SCORE CALCULATION
USING ARTERIAL CALCIFICATION HEATMAPS AS FEATURE WEIGHTING.
Model
Input
Original
Acc(%)
Weighting
Acc(%)
Concatenation
Acc(%)
Resnet18
66.67
72.46
75.00
Resnet18+SimAM
64.67
76.27
77.56
Resnet18+CBAM
65.33
76.69
79.24
Resnet18+SA
69.00
77.12
78.73
Furthermore, we compared and adjusted our model's
output to calculate AAC scores in the AAC-24 point, adopting
the same classification standards as [11], categorizing each
patient into low, medium, and high levels according to their
AAC scores. Table IV shows the accuracy analysis of AAC
score calculation using the same scoring method as [11]. Our
method achieved an accuracy of 86.66%, surpassing the
average accuracy rate of 80.08% reported in [11]. The grading
formula used in the paper, where represents the patient's
AAC score, is as follows:
󰇛󰇜
󰇛󰇜 
󰇛󰇜 
󰇛󰇜 
International Journal of Applied Sciences & Development
DOI: 10.37394/232029.2024.3.17
Chung-Kuan Wu, Che-Yu Chiang, Jun-Wei Hsieh
E-ISSN: 2945-0454
174
Volume 3, 2024
TABLE IV. ANALYSIS OF ACC SCORE CALCULATION ACCURACY
USING THE SAME SCORING METHOD AS DESCRIBED IN [11].
Method
Acc(%)
[11]
80.08
our method
86.66
The AAC grade is an essential indicator of vascular
calcification (VC) and is closely linked to vascular diseases.
We have developed a CNN-based attention model for
classifying AAC grades, incorporating various attention
blocks to amplify subtle features. Compared to conventional
feature extraction backbone networks, our method achieves
precise classification of AAC grades. We believe this model
can aid physicians in reducing the time needed for
interpretation and enhancing the accuracy of their assessments,
thereby enabling more objective treatment planning.
[1] J. Bover et al., “Clinical approach to vascular calcification in patients
with non-dialysis dependent chronic kidney disease: Mineral-bone
disorder-related aspects,” Frontiers in Medicine, vol. 8, May 2021.
[2] S.-C. Ren et al., “Vascular calcification in chronic kidney disease: An
update and perspective,” Aging and disease, vol. 13, no. 3, p. 673, Jun.
2022.
[3] Kauppila, L. I., et al. “New indices to classify location, severity and
progression of calcific lesions in the abdominal aorta: a 25-year follow-
up study,” Atherosclerosis, vol. 132, no. 2, pp. 245–250, 1997.
[4] Schousboe, J. T., Wilson,K. E., and Kiel,D. P., “Detection of
abdominal aortic calcification with lateral spine imaging using DXA,”
Journal of Clinical Densitometry, vol. 9, no. 3, pp. 302–308, Jul. 2006.
[5] Chen, P.-Y., Chang, M.-C., Hsieh, J.-W., and Chen, Y.-S., “Parallel
residual bi-fusion feature pyramid network for accurate single-shot
object detection,” IEEE Transactions on Image Processing, vol. 30, pp.
9099-9111, 2021.
[6] Bochkovskiy, Alexey, Chien-Yao Wang, and Hong-Yuan Mark Liao.
"Yolov4: Optimal speed and accuracy of object detection." arXiv
preprint arXiv:2004.10934, 2020
[7] He, K., Zhang, X., Ren, S., and Sun, J., “Deep residual learning for
image recognition,” Proceedings of the IEEE conference on computer
vision and pattern recognition, 2016.
[8] Woo, S., Park, J., Lee, J.-Y., and Kweon, I. S., “Cbam: Convolutional
block attention module,” Proceedings of the European conference on
computer vision (ECCV), pp.3-19, 2018.
[9] Yang, L., et al., “SimAM: A Simple, Parameter-Free Attention Module
for Convolutional Neural Networks,” in Proceedings of Machine
Learning Research, vol. 139, pp.11863-11874, 2021.
[10] Zhang, Q., and Yang, Y., “SA-Net: Shuffle Attention for Deep
Convolutional Neural Networks,” IEEE International Conference on
Acoustics, Speech and Signal Processing, 2235-2239, 2021.
[11] Sharif, N., et al., “Machine learning for abdominal aortic calcification
assessment from bone density machine-derived lateral spine
images,” EBioMedicine, vol. 94, 104676, July 2023.
5. Conclusion
References
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare
that are relevant to the content of this article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
International Journal of Applied Sciences & Development
DOI: 10.37394/232029.2024.3.17
Chung-Kuan Wu, Che-Yu Chiang, Jun-Wei Hsieh
E-ISSN: 2945-0454
175
Volume 3, 2024