Use of artificial intelligence in the diagnosis of arterial calcification

Full Text

BACKGROUND

The incidence of circulatory system diseases in the Russian Federation has been steadily increasing in the last two decades, growing 2.047 times between 2000 and 2019 [1]. Vascular calcification involves the deposition of calcium salts in the arterial wall, which leads to vascular wall remodeling [2]. Intimal calcification is focal and associated with atherosclerosis, whereas medial calcification is diffuse and involved in the pathogenesis of conditions such as diabetes mellitus, peripheral artery disease, and chronic kidney disease [3]. X-ray imaging is the gold standard for diagnosing vascular calcifications [4]. However, because of increasing data and the need to reduce the time for diagnosing, work efficiency inevitably decreases [5]. These circumstances necessitate the search for innovative ways to improve the quality of work of imaging specialists.

The active development and introduction of artificial intelligence (AI) into clinical practice have helped specialists address these issues. According to the available literature, until recently, AI was used for X-ray diagnosis of five types of vascular calcifications: coronary artery calcification (CAC), thoracic aorta calcification (TAC), abdominal aorta calcification (AAC), carotid artery calcification (CaAC), and mammary artery calcification (MAC).

This study aimed to analyze the national and international literature on the use of AI in the diagnosis of various vascular calcifications, summarize the prognostic value of vascular calcification, and evaluate aspects that hinder the diagnosis of vascular calcification without using AI.

MATERIALS AND METHODS

A literature search was performed in PubMed, Web of Science, Google Scholar, and eLibrary. The search was conducted using the following keywords: artificial intelligence, machine learning, vascular calcification, and their equivalents in Russian. The search covered the period from database inception till July 2023. The authors reviewed independently the titles and abstracts of the articles and retrieved the full text of relevant publications. Moreover, the reference lists of relevant studies were reviewed.

RESULTS

The studies included in the review compared the diagnostic abilities of clinicians and AI using the same images and subsequently assessed the accuracy, speed, and other parameters.

The sites of vascular calcification vary greatly, resulting in differences in their prognostic value. Table 1 summarizes the prognostic value of vascular calcifications depending on the location.

 

Table 1. Prognostic value of vascular calcifications depending on the location

Vascular calcification type	Prognostic value
Coronary artery calcification	– Marker of the severity of coronary artery atherosclerosis – Cardiovascular risk assessment according to the Framingham risk score – CAD predictor – Marker of chemotherapy-induced cardiotoxicity in patients with cancer
Thoracic aorta calcification	– Marker of increased CAD risk – Detection of an increased risk of ischemic stroke – Embolism risk detection
Abdominal aorta calcification	– Detection of obstructive CAD – Prediction of asymptomatic CAD – Marker of congestive heart failure
Carotid artery calcification	– Marker of atherosclerosis of the head and neck blood vessels – Stroke risk detection – Prediction of the risk of cerebrovascular adverse events in young people
Mammary artery calcification	– Detection of the risk of cardiovascular events in women – Association with chronic kidney disease, diabetes mellitus, and bone diseases

Note: CAD, coronary artery disease.

 

Coronary artery calcification

The Framingham risk score is a tool for cardiovascular risk assessment, which includes the assessment of risk factors such as age, sex, and blood pressure [7]. However, a large-scale prospective study with 7 years of follow-up found that CACs detected using computed tomography (CT) can improve the risk prognosis obtained using the Framingham risk score alone. In a study with 7.6 years of follow-up, M.H. Criqui et al. demonstrated a good prognostic value of the assessment of CAC severity based on the vascular volume and density [8]. The risk of coronary artery disease (CAD) positively correlated with the CAC volume and negatively correlated with the CAC density [8].

The CAC severity was assessed by multiplying the density of calcified plaques by the area of calcification. The overall CAC was the sum of the results calculated at each level. Generally, older patients represent a major risk group for CAD [1]. However, a 12.5-year study revealed an increased risk for CAD and death among individuals aged 32–46 years, even in cases of mild CAC [9]. These findings suggest that CAC-related information has a high prognostic value for detecting cardiovascular risk in nearly all age groups. Table 2 shows the relationship between CAC severity and risk for adverse cardiovascular events.

 

Table 2. Relationship between the severity of coronary artery calcification and the risk for adverse cardiovascular events

Coronary artery calcification assessment	Calcification risk
0	None
1–10	Low
11–100	Moderate
101–400	Moderate to high
>401	High

 

Thoracic aorta calcification

TAC is commonly found in patients with hypertension [10]. Furthermore, recurrent evidence has linked TAC to an increased risk for CAD and death [10, 11]. In a study of 2,618 patients, Y. Itani et al. found that TAC was efficient in ischemic stroke risk assessment [12]. In a study of patients with indications for cardiovascular surgery, R. Lee et al. found that preoperative CT screening for TAC can identify high-risk areas and reduce the risk of aortic embolism and stroke [13]. Thus, TAC severity can be used not only to predict the risk of cardiovascular accidents but also to detect cerebrovascular changes.

Abdominal aorta calcification

A study of 58 patients revealed that CT-AAC assessed AAC correlated with CAC severity. In turn, the absence of AAC made it possible to rule out CAD [14]. Moreover, AAC can be used as an additional tool for detecting asymptomatic CAD and an independent risk factor for congestive heart failure [15, 16]. AAC has a significant prognostic value for the skeletal system. Y.Z. Bagger et al. analyzed 2,662 healthy postmenopausal women and reported that AAC correlated with an increased risk of osteoporosis of the proximal femur [17]. Moreover, in a study of 5,994 men aged 65 years, P. Szulc et al. confirmed the correlation between AAC and an increased risk for femoral fracture in older men [18].

Carotid artery calcification

CaAC is an important predictor of cerebrovascular diseases [19]. Intracranial internal carotid artery calcification (ICAC) is an important marker of intracranial hypertension in patients of various ethnicities and strongly correlated with the risk for stroke [19–21]. A study of approximately 2,000 patients revealed that ICAC was common in young people. However, whether ICAC at a young age is similar to that at an older age is unclear and, therefore, may increase the risk of stroke later in life [22].

Mammary artery calcification

Z. Huang et al. analyzed 213 female patients and revealed that MAC correlated with CAC and CAD severity [23]. EV Bochkareva et al. analyzed 4,274 digital mammograms of women aged 40–93 years and found a significant and strong correlation between age and MAC [24]. E.V. Bochkareva et al. also demonstrated that MAC was associated with chronic kidney disease, diabetes mellitus, cerebrovascular diseases, and low bone mineral density [25–27]. Unfortunately, female patients are often unaware of cardiovascular diseases, which pose one of the most serious threats to women’s health [28]. Thus, considering the availability of MAC assessment and its diagnostic value in assessing cardiovascular risk in women, imaging specialists should pay close attention to this issue.

Difficulties in imaging diagnosis of vascular calcifications

An analysis of the evolution of X-ray diagnostic services in Russia in 2014–2019 showed that the number of medical images is increasing annually, and imaging specialists have to interpret images every 3–4 s during an 8-h working day [29]. A study of the physical condition of 40 imaging specialists before and after a working day revealed that after working for 1 day, their ability to concentrate decreases dramatically, whereas symptoms of asthenopathy increase [30]. According to available data, 75% of claims concerning the low quality of medical care provided by imaging specialists are related to diagnostic errors [31].

Accurate assessment of the severity of vascular calcifications is difficult. The shape of the calcified foci is variable, and deviations are common. For example, various modifications of CT scans have been used to diagnose CAC [32]; however, they require additional equipment, increasing the economic burden on healthcare facilities and the patient’s radiation exposure (which can be reduced if the diagnosis is made by standard methods). Moreover, radiologists do not always assess vascular calcifications detected on CT scans, which could be used for indirect assessment of coronary calcification. Nonetheless, the gold standard is CT with cardiac synchronization and a specified examination area (field of view) [33, 34].

The role of artificial intelligence in the assessment of arterial calcification

CACs

The first automatic assessment of CAC severity using AI was performed in 2007. For each candidate, 64 features were created, and nearest-neighbor clustering was applied. This metric algorithm for automatic object classification or regression had an accuracy of 73.8% [35]. Since then, various approaches to feature development, including spatial and geometric characteristics, have been actively studied [36–38]. Because selecting objects on noncontrast CT scans is technically challenging, recording information from CT scans with the determination of coronary calcification has become a common strategy [39–41]. To assess coronary calcification, electrocardiographic (ECG) synchronization is commonly used to capture images in the diastolic phase, followed by image reconstruction by stitching. A support vector machine-based algorithm achieved a sensitivity of 98.9% and a prognostic value of 94.8% [42]. Such machine-learning (ML) algorithms were actively used until 2016; however, their use was challenging because of the need for manual control [42].

To further improve efficacy, an artificial neural network (ANN) with deep learning (DL) function was selected as the primary option [43]. DL initially demonstrated low efficacy; however, continuous improvement of the ANN allowed for increased efficacy and accurate, automatic scoring [44–51].

In a study on CAC assessment, B.D. de Vos et al. discovered that the results obtained using DL were nearly identical to those calculated manually. The Agatston score (gold standard of CT calcium scoring in clinical practice) was determined in <0.3 s [52].

The U-Net algorithm is an extension of the ANN intended for more efficient learning requiring fewer resources [53]. N. Gogin et al. confirmed the efficacy of DL based on the U-Net architecture, which proved to be extremely close to the performance of other algorithms [46]. U-Net classified the risks correctly in 86% of cases. Notably, U-Net allows uploading images directly without losing pixel information [49].

CT with cardiac synchronization is less widely used globally than CT without ECG [54]. If the reliability of CAC information obtained without ECG increases, the number of CT scans can be reduced, thus decreasing economic expenditures and radiation exposure [55]. Numerous interferences and artifacts significantly lower the accuracy of manual CAC assessment using CT without ECG [56].

AI has played a significant role in CAC diagnosis using CT without ECG. More than a decade ago, I. Isgum et al. found that ML algorithms ensure the Agatston score assessment with an accuracy of 82.2% when using low-dose chest CT [57]. Many researchers have attempted to level out interference and artifacts and reduce the effect of noncalcified components (stents) on the diagnosis. Their developed DL algorithms allowed for increasing the diagnostic value of low-quality images [58, 59]. According to Z. Sun et al., DL algorithms improved the signal-to-noise ratio of low-dose CT by 27.7% and increased the specificity of CAC detection by 41% by eliminating artifacts [60]. Notably, the rate of correct interpretation of CT without ECG was only 70%; however, the correlation coefficient between the results obtained using AI and manual analysis was 0.923 [56].

B. Yacoub et al. revealed that the sensitivity, specificity, and area under the curve (AUC) of CAC detection on noncontrasted chest CT scans using AI were superior to those of manual assessment. This suggests that AI can be superior to human assessment when analyzing CT findings without ECG [34]. In one study, the efficacy of AI in analyzing CT findings without ECG was confirmed at four study sites. All study samples showed high sensitivity and good prognostic value, which increases the quality of the results [61]. The use of AI in diagnosing CAC based on CT findings without ECG is currently considered a reliable method of assessing data.

Notably, researchers have begun to use AI capabilities to diagnose CAC using other devices, such as when assessing chest X-ray images [62]. P.I. Kamel et al. developed a classification of total calcium parameters on chest X-ray images using deep ANN, making it possible to reduce the need for CT in some patients [63]. In this study, the AUC reached 0.73 and 0.7 in the anteroposterior and lateral images, respectively, when detecting CAC. Moreover, a study proposed a neural network that could analyze invasive coronary angiography images within seconds and detect CAC with a F1 of 0.802 [64].

Thoracic aorta calcification

In contrast to CAC, the accuracy of TAC detection on CT scans does not depend on the heart contraction intensity [65]. For example, in a study by I. Isgum et al., the TAC detection rate was 97.9% when using the k-nearest neighbor method, which correlated with manual findings [65]. In recent years, DL made it possible to detect CAC and TAC simultaneously [66]. Similarly, SGM van Velzen et al. used these methods with various CT modalities (including CT with the determination of coronary calcium, low-dose chest CT, and positron-emission CT), with an intraclass correlation coefficient of 0.68–0.98 [67]. Notably, a sensitivity of 98.4% was observed in a study using a convolutional neural network (CNN) for TAC detection, which made it possible not only to detect TAC in the ascending and descending aorta and the aortic arch but also to assess the risk level [68]. Thus, radiologists now routinely employ AI to automatically assess CAC and TAC.

Abdominal aorta calcification

DL advancements made it possible to automatically detect AAC, which was confirmed in two studies [69, 70]. Dual-energy X-ray absorptiometry is a diagnostic tool that assesses the risk for fractures. The increasingly wide use of this method suggests the possibility of automating AAC assessment; however, because of technical difficulties, it is not routinely used for AAC detection. S. Reid et al. classified AAC using a densitometry-based CNN, with a high degree of agreement with manually obtained findings; the Kappa index was 0.71 [69]. CT provides obvious advantages over dual-energy X-ray absorptiometry for the qualitative assessment of aortic calcification. P.M. Graffy et al. effectively implemented automatic detection of AAC using abdominal CT in >9,000 patients; the authors attributed their success to the use of mask region-based CNN [70]. This study also assessed the prevalence of AAC based on quantitative data, which supports the importance of AI.

In summary, AI can be used to automatically quantify AAC; however, available data are extremely limited.

Carotid artery calcification

CT in combination with AI can be used to detect calcification of both extracranial and intracranial internal carotid arteries [71, 72, 73]. G. Bortsova et al. used four DL networks with a structure similar to that of U-Net. The accuracy of ICAC detection was higher than manual assessment, with a sensitivity of 83.8% and a prognostic value of 88% [73]. Manual assessment of CIC necessitates meticulous analysis; moreover, it is prone to errors, with similar probability (e.g., bone calcification). Given this, the excellent accuracy of DL is of great importance.

Magnetic resonance imaging (MRI) can detect the most significant differences between CaAC and other vascular calcifications [71]. In earlier studies, MRI showed low accuracy in detecting calcifications [71]. However, the use of simultaneous noncontrast angiography and intraplaque hemorrhage (SNAP) has improved the ability to detect calcifications using MRI. SNAP inverts all signals by pulse inversion, followed by T1-weighted inversion recovery and a proton density-weighted control scan with dual gradient echo, providing high-quality images of the cranial and cervical spine arteries [74]. Although SNAP effectively detects calcification, it is prone to motion artifacts and has a long acquisition time. Goal-SNAP and quick SNAP can be used to address this issue [75, 76]. In this study, ML algorithms such as random forest are similar to ANNs in terms of calcification detection; however, DL may be more effective in segmenting vascular components, which requires further research.

Mammary artery calcification

MAC is visualized on mammograms; however, calcifications have extremely diverse manifestations. They can be bifurcated, overlapping, or truncated, with varying intensities [77]. As a result, MAC is difficult to quantify manually because of its heterogeneous presentations.

Mammography is an X-ray diagnosis method used for screening and intended for breast cancer detection. It has been included in the scope of preventive medical examination since 2012. In line with the regulations of the Ministry of Health of the Russian Federation, the frequency of this study increases annually.

Several authors have considered the use of DL in this context. Mammograms were divided into sections because the amount of data was too large for direct entry [53]. The 12-layer CNN defines MAC detection as a second-order task [28]. Although it successfully distinguished the presence and absence of MAC, the accuracy of the quantitative assessment was insufficient [28]. Furthermore, data analysis and processing were time-consuming because of the need to process each segment separately. Subsequently, researchers improved the CNN considering the described drawbacks, suggesting the use of a simple contextual U-Net (SCU-Net) and a dense U-Net (DU-Net) [33, 53]. SCU-Net is a simpler version of U-Net that addresses MAC accounting for <1% of the images, resulting in a significant amount of data and preventing the system’s efficient training. DU-Net eliminates this problem by considerably improving the efficacy of the CNN, with an accuracy of 91.47% and a sensitivity of 91.22% [53].

In summary, DL used for automatic MAC detection has advanced considerably; however, no public dataset is available for the unification in this field, which necessitates further studies.

DISCUSSION

AI can facilitate and improve the work of imaging specialists regarding vascular calcification by performing preliminary screening and enhancing data processing efficacy.

The accuracy of AI in X-ray diagnosis was determined by AI algorithms and image characteristics. The efficacy of ML and DL algorithms has improved significantly, increasing the diagnostic value of these techniques. As imaging technology evolves, image quality improves, which intensifies the efficiency of diagnosis. AI capabilities depend on the high quality of databases required for training, which can also be useful in creating public databases or testing platforms. Thus, engineers and physicians should work to improve AI diagnostic capabilities based on AI algorithms and image quality.

Although AI has demonstrated good results in five types of vascular calcifications, it also has potential diagnostic value in renal artery calcification, a condition with confirmed prognostic value for hypertension and proteinuria [79]. Moreover, the number of studies on AAC and CaAC is modest, which increases interest in their research. Table 3 provides a comparison of the efficacy of AI in the diagnosis of each type of vascular calcification.

 

Table 3. Comparison of the efficacy of AI in the diagnosis of five types of vascular calcification

Vascular calcification type	Number of studies	Use of X-ray imaging	Use of CT	Use of MRI	Use of DL
Coronary artery calcification	Large	Yes	Yes	No	Pixel-based and end-to-end DL in combination with GAN reduces the level of noise and artifacts for a reliable automatic assessment of CAC and interpretation of the results
Thoracic aorta calcification	Moderate	No	Yes	No	DL ensures the automatic detection of TAC and CaAC and enables the assessment of the findings in various portions of the aorta
Abdominal aorta calcification	Small, more studies are required	Yes	Yes	No	DL techniques such as the mask region-CNN enable accurate qualitative assessment of AAC when using DEXA and CT
Carotid artery calcification	Small, more studies are required	No	Yes	Yes	The efficacy of DL is similar to that of ML and superior to human performance, with the added possibility of using MRI
Mammary artery calcification	Moderate	Yes	No	No	U-Net, SCU-Net, DU-Net, and other improved DL systems facilitate MAC detection

Note: AAC, abdominal aorta calcification; CAC, coronary artery calcification; CT, computed tomography DEXA, dual-energy X-ray absorptiometry; DL, deep learning; DU-Net, dense U-Net: GAN, generative adversarial network; MAC, mammary artery calcification; ML, machine learning; MRI, magnetic resonance imaging; Mask region-CNN, mask region-based convolutional neural network; SCU-Net, simple contextual U-Net; TAC, thoracic aorta calcification.

 

CONCLUSION

Artificial intelligence has proven to be effective in the diagnosis of vascular calcifications. In addition to improved accuracy and efficiency, its level of detail is superior to that of manual diagnosis methods. AI has advanced to the point that imaging specialists can automatically detect vascular calcification. Artificial intelligence can contribute to the successful development of X-ray imaging in the future.

ADDITIONAL INFORMATION

Funding source. This study was not supported by any external sources of funding.

Competing interests. The authors declare that they have no competing interests.

Authors’ contribution. All authors made a substantial contribution to the conception of the work, acquisition, analysis, interpretation of data for the work, drafting and revising the work, final approval of the version to be published and agree to be accountable for all aspects of the work. Yu.A. Trusov, V.S. Chupakhina — development of the concept and design of the study, scientific revision of the manuscript; A.S. Nurkaeva — data analysis and search, writing the manuscript; N.A. Yakovenko — data analysis, writing the manuscript; I.V. Ablenina — data analysis, editing the text of the manuscript; R.F. Latypova — data analysis, verification and approval of the text; A.P. Pitke — writing a manuscript, obtaining factual data; A.A. Yazovskikh — editing the text of the article, analyzing data; A.S. Ivanov — obtaining data, writing the text of the article; D.S. Bogatyreva — approval of the final version of the manuscript, editing of the text of the article; U.A. Popova — data analysis, participation in the writing of the article; A.F. Yuzlekbayev — analysis and data acquisition.

About the authors

Yuri А. Trusov

Samara State Medical University

Email: yu.a.trusov@samsmu.ru
ORCID iD: 0000-0001-6407-3880
SPIN-code: 3203-5314
Russian Federation, Samara

Victoria S. Chupakhina

Rostov State Medical University

Email: chupalhina@bk.ru
ORCID iD: 0009-0003-8318-3673
SPIN-code: 4402-7476
Russian Federation, Rostov-on-Don

Adilya S. Nurkaeva

Bashkir State Medical University

Email: vkomissiya@inbox.ru
ORCID iD: 0009-0006-8621-5580
SPIN-code: 3307-5546
Russian Federation, Ufa

Natalia A. Yakovenko

Sechenov First Moscow State Medical University

Email: tigris2011@yandex.ru
ORCID iD: 0009-0005-6726-9623
SPIN-code: 4415-2236
Russian Federation, Moscow

Irina V. Ablenina

Orenburg State Medical University

Email: aninelba@gmail.com
ORCID iD: 0009-0006-6222-9339
SPIN-code: 4123-3336
Russian Federation, Orenburg

Roksana F. Latypova

Orenburg State Medical University

Email: roxevansss@gmail.com
ORCID iD: 0009-0004-5057-6451
SPIN-code: 3542-3376
Russian Federation, Orenburg

Aleksandra P. Pitke

Orenburg State Medical University

Email: pitkea00@gmail.com
ORCID iD: 0009-0002-1111-759X
SPIN-code: 3726-4213
Russian Federation, Orenburg

Anastasiya A. Yazovskih

Bashkir State Medical University

Email: anyaz.bgmu@yandex.ru
ORCID iD: 0000-0002-3955-0830
SPIN-code: 3543-5323
Russian Federation, Ufa

Artem S. Ivanov

Bashkir State Medical University

Email: artem.ivanov656@yandex.ru
ORCID iD: 0009-0000-3562-8293
SPIN-code: 4834-5324
Russian Federation, Ufa

Darya S. Bogatyreva

Russian National Research Medical University named after N.I. Pirogov

Email: diria1012@yandex.ru
ORCID iD: 0009-0004-5055-8819
SPIN-code: 3331-3421
Russian Federation, Moscow

Ulyana A. Popova

Russian University of Medicine

Email: ulyanka.popova.2000@gmail.com
ORCID iD: 0009-0002-7994-5631
SPIN-code: 3452-2543
Russian Federation, Moscow

Azat F. Yuzlekbaev

Bashkir State Medical University

Author for correspondence.
Email: ztl5@rambler.ru
ORCID iD: 0009-0002-8799-4732
SPIN-code: 4812-3213
Russian Federation, Ufa

References

Supplementary files