Artificial intelligence in coronary computed tomography angiography

Abstract

Coronary computed tomography angiography (CCTA) is recommended as a frontline diagnostic tool in the non-invasive assessment of patients with suspected coronary artery disease (CAD) and cardiovascular risk stratification. To date, artificial intelligence (AI) techniques have brought major changes in the way that we make individualized decisions for patients with CAD. Applications of AI in CCTA have produced improvements in many aspects, including assessment of stenosis degree, determination of plaque type, identification of high-risk plaque, quantification of coronary artery calcium score, diagnosis of myocardial infarction, estimation of computed tomography-derived fractional flow reserve, left ventricular myocardium analysis, perivascular adipose tissue analysis, prognosis of CAD, and so on. The purpose of this review is to provide a comprehensive overview of current status of AI in CCTA.

Key Words: Coronary computed tomography angiography; Coronary artery disease; Artificial intelligence; Deep learning; Machine learning; Prognosis

Core Tip: The application of artificial intelligence in coronary computed tomography angiography mainly focuses on the following aspects: (1) Studies based on the coronary arteries and plaques for determination of stenosis degree, identification of plaque types, quantification of coronary artery calcium score, prediction of myocardial infarction, and prognosis evaluation; (2) Studies around the perivascular adipose tissue, which were mainly conducted using radiomics analysis and machine learning algorithm, for improvement of risk stratification; and (3) Studies based on the texture analysis of the left ventricular myocardium for assessment of functionally significant stenosis or for prognosis evaluation.

INTRODUCTION

Coronary computed tomography angiography (CCTA) has merged as a first-line diagnostic tool in the non-invasive evaluation of patients with suspected coronary artery disease (CAD), as recommended in the international guidelines[1,2]. With rich information provided in the luminal stenosis, the morphology and composition of plaques, and the overall circulation, CCTA can safely rule out the obstructive CAD and improve prognosis.

However, the information derived from CCTA images is recognized and interpreted by human readers, and varies among different scanning protocols, scanners, contrast medium injection protocols, and readers. The arrival of artificial intelligence (AI) brought hope that it can be applied for intelligent decision-making with autonomous acquired knowledge by identifying and extracting patterns among a group of observations[3,4].

With the frontline role of CCTA in the diagnostic strategies for CAD, “big data” is available and offers an optimal platform to bridge AI with CCTA. Recently, AI techniques in CCTA have gained much attention and have been widely applied in clinical care ranging from diagnosis to prognostic stratification. We seek to summarize the recent application of AI techniques in CCTA images, so as to investigate and identify the most important and promising research topics, the problems that have been resolved and remain to be resolved, and the future directions with many challenges and opportunities.

CURRENT APPLICATION OF AI IN CCTA

The application of AI in CCTA images mainly focuses on the following aspects: (1) Studies based on the coronary arteries and plaques for determination of stenosis degree, identification of plaque types, quantification of coronary artery calcium (CAC) score, prediction of myocardial infarction (MI), and prognosis evaluation; (2) studies around the perivascular adipose tissue (PVAT), which were mainly conducted using radiomics analysis and machine learning (ML) algorithm, for improvement of risk stratification; and (3) studies based on the texture analysis of the left ventricular myocardium (LVM) for assessment of functionally significant stenosis or for prognosis evaluation, as shown in Figure 1.

Figure 1 The application of artificial intelligence in coronary computed tomography angiography.

AUTOMATIC DETECTION AND CLASSIFICATION OF CORONARY ARTERY PLAQUE AND STENOSIS

Since different grades of coronary artery stenosis and varying types of plaque would lead to different patient management strategies, it is therefore crucial to: (1) Detect and determine the stenosis; (2) Detailedly characterize plaques (i.e., non-calcified, calcified, mixed plaques); and (3) Identify the so-called “high-risk” plaque features. Recently, there are already applications of AI techniques in related CCTA fields, including stenosis evaluation and plaque characterization. Commonly, the anatomical evaluation of coronary stenosis and quantification of plaques rely on a relative accurate segmentation and successful automatic lesion localization in CCTA images. Several vendors are developing AI-based platform for stenosis evaluation. However, the identification of “high-risk” plaques remains challenging, and only a few studies have been proposed but are of great promise with prognostic value.

Kang et al[5] proposed a structured learning technique for automatic detection of obstructive and non-obstructive CAD on CCTA. Taking the visual identification of lesions with stenosis ≥ 25% by three expert readers, using consensus reading, as the reference standard, the method achieved a high sensitivity (93%), specificity (95%), and diagnostic accuracy (94%), with an area under the curve (AUC) of 0.94. Zreik et al[6] employed a multi-task recurrent convolutional neural network to determine the stenosis severity based on the MPR view of a coronary artery extracted from the CCTA scan, as well as to automatically detect and characterize the coronary plaques. The approach achieved an accuracy of 0.80 for the determination of the anatomical significance of the coronary artery stenosis, and 0.77 for the detection and characterization of coronary plaques. Wei et al[7] developed a topological soft-gradient (TSG) detection method to prescreen for noncalcified plaque (NCP) candidates, which achieved AUCs of 0.87 ± 0.01 and 0.85 ± 0.01 in the training and validation sets, respectively. Jawaid et al[8] utilized support vector machine algorithms for automated detection of NCPs, and their approach achieved a detection accuracy of 88.4% with respect to the manual expert and a dice similarity coefficient of 83.2%.

In 2017, Kolossváry et al[9] investigated whether radiomics analysis improves the identification of coronary plaques with or without Napkin-ring sign (NRS). NRS is characterized as a so-called “high-risk” plaque features, which is defined as a plaque core with low CT attenuation apparently in contact with the lumen that is surrounded by a ring-shaped higher attenuation as napkin ring like in CCTA images[10,11]. However, the identification of the NRS remains challenging because it is assessed by a qualitative read of CCTA images which is affected by clinical experience and intra-/inter-reader variability[12]. Based on the segmented CCTA datasets, 8 conventional quantitative metrics and 4440 radiomic features were extracted. They found that none of the conventional quantitative parameters but 20.6% (916/4440) of radiomics features were significantly different between NRS and non-NRS plaques (Bonferroni-corrected P < 0.0012). In addition, almost half of the features (418/916) reached an AUC > 0.80, of which three features, including short- and long-run low gray-level emphasis and surface ratio of high attenuation voxels to total surface, exhibited excellent discriminatory value with AUCs of 0.918, 0.894, and 0.890, respectively. In 2019, the same research group validated the radiomics features extracted from CCTA in an ex-vivo histological study. One ML algorithm incorporating 13 parameters was superior compared with visual assessment (AUC = 0.73 vs 0.65) in the identification of advanced lesions[13].

DEEP LEARNING FOR AUTOMATIC CAC SCORING

CAC scoring plays a key role in risk stratification of CAD. Non-contrast-enhanced cardiac CT, which is routinely acquired as a stand-alone test or an adjunct study prior to CCTA, is considered as the reference for quantification of CAC. CAC is defined as a high-attenuation area with > 130 HU in at least three contiguous pixels in non-contrast-enhanced cardiac CT. Recently, it has been shown that CAC can be also detected in CCTA images, which could reduce the radiation dose of a typical cardiac CT examination by 40%-50%[14]. Besides, the increased visibility of the coronary arteries in CCTA compared to non-contrast-enhanced cardiac CT could improve the identification of CAC. However, manual quantification of CAC requires substantial clinical experience to identify and make of every calcified lesion in each image slice, which is a time-consuming process. Consequently, a series of automatic methods have been proposed for CAC scoring in CCTA. Many investigations have shown promising results for clinical application in this field.

Some researchers[15,16] developed the automatic methods using two stages, including: (1) Segmentation of the coronary arteries; and (2) Identification of the CAC with the deviation from a trend line through the lumen intensity, or the voxels above a specific HU threshold, or the deviation from a model of non-calcified artery segments.

Wolterink et al[17,18] proposed an automatic CAC quantification method without a need for segmentation of the coronary artery tree in CCTA images using a combination of a convolutional neural network (CNN) and a Random Forest classifier. Thereafter, the same working group further extended and optimized their framework using a pair of CNNs in five ways[18], and the automatic CAC scoring in CCTA using a pair of CNNs yielded a high correlation (Pearson P = 0.950) and high consistency (intraclass correlation coefficient of 0.944) with the reference CAC scoring in non-contrast-enhanced CT.

In 2020, Fischer et al[19] proposed a novel fully automated algorithm using recurrent neural network with long short-term memory to detect CAC from CCTA data in a total of 565 vessels. An accuracy of 90.3% [95% confidence interval (CI): 88.0%-90.0%] was achieved on a per-vessel basis.

In summary, the CAC scoring performed on routine CCTA images without additional radiation exposure is highly desirable and the application of AI has provided considerable progress in the field and would become more influential in the clinical setting. In the near future, with the widespread application of AI techniques, CAC scoring using CCTA may eliminate the need for separate dedicated coronary calcium-scoring non-contrast enhanced CT scans.

IDENTIFICATION OF MYOCARDIAL ISCHEMIA

ML-based fractional flow reserve-CT for detection of functionally significant stenosis

It has been demonstrated that the anatomically significant appearance of a coronary stenosis is insufficient to detect hemodynamic significance and does not always equate with functional significance, which is particularly true for intermediate type coronary lesions[20,21]. Fractional flow reserve (FFR) performed during cardiac catheterization has been the reference standard in the detection of lesion-specific ischemia and is recommended for therapeutic decision-making[22]. However, the invasive measurement with a pressure wire and the relatively high cost restrict the clinical application of FFR.

Recently, novel non-invasive approaches utilizing ML algorithms for determination of FFR based on conventional CCTA images (FFR-CT) were developed and validated with a considerable diagnostic accuracy. The most popular algorithm is FFR-CT_ML (Figure 2). FFR-CT_ML was developed by Itu et al[23] in 2016 and provided by only one vendor (Siemens Healthineers, Germany) for research purpose. With the rapid development of AI, some FFR-CT platforms were provided for commercial use, such as the DEEPVESSE-FFR Platform provided by Keya Medical (Beijing, China). The DEEPVESSE-FFR Platform was developed by Wang et al[24] using MLNN + BRNN and has been commercially available since 2020.

Figure 2 The workflow of the fractional flow reserve-computed tomography derivation. ¹A total of 12000 coronary anatomies were generated; ²twenty-eight geometric features were extracted from the synthetically generated database; ³a deep neural network with four hidden layers was used to train the machine learning-based model. FFR-CT: Fractional flow reserve-computed tomography; CCTA: Coronary computed tomography angiography.

So far, ML-based FFR-CT has been evaluated in several multi-center and single-center studies[23-35] using a threshold of ≤ 0.80 acquired from invasive FFR to detect lesion-specific ischemia. It has been demonstrated that ML-based FFR-CT performed equally in detecting flow-limiting stenosis compared with the computer fluid dynamics (CFD) based FFR-CT (FFR-CT_CFD)[26], while the FFR-CT_CFD algorithm is time-consuming and heavily affected by the image quality[25,27,36]. The performance of ML-based FFR-CT in the related literature is summarized in Table 1.

Table 1 Summary of the current literature on machine learning-based fractional flow reserve-computed tomography.

Ref.	Journal	Prospective	Multi- or single center	Platform	No. of patients	No. of vessels	Compared with CT-FFRCFD	Accuracy	AUC
Itu et al[23], 2016	Journal Application Physiology	No	Single center	-	87	125	Yes	Per-lesion: 83%	Per-lesion: 0.90
Coenen et al[25], 2018	Circulation: Cardiovascular Imaging	Yes	The MACHINE registry	cFFR, version 2.1, Siemens	351	525	Yes	Per-lesion: 78%Per-patient: 85%	Per-lesion: 0.84
Tesche et al[26], 2018	Radiology	No	Single Center	cFFR, version 1.4, Siemens	85	104	Yes	Per-lesion: 88%; Per-patient: 92%	Per-lesion: 0.89; Per-patient: 0.91
Mastrodicasa et al[34], 2019	Journal of Cardiovascular Computed Tomograph	No	Single center	cFFR, version 3.0, Siemens	10/40	160	No	IRIS: 82%; FBP: 82%	-
Baumann et al[32], 2019	European Journal of Radiology	No	The MACHINE registry	cFFR, version 2.1, Siemens	351	525	No	-	Per-patient: Women:0.83; Men: 0.83
Doeberitz et al[27], 2019	European Radiology	No	Single center	cFFR, version 2.1, Siemens	48	103	No	-	Per-lesion: 0.93
Wang et al[24], 2019	Journal of Geriatric Cardiology	Yes	Single center	DEEPVESSE-FFR Platform	63	71	No	Per-lesion: 89%; Per-patient: 87%	Per-lesion: 0.93; Per-patient: 0.93
Tesche et al[30], 2020	Journals of the American College of Cardiology: Cardiovascular Imaging	Yes	The MACHINE registry	cFFR, version 2.1, Siemens	314	482	No	Per-lesion: 78%; CAC ≥ 400: 76%CAC 0-100: 79%; CAC 100-400: 76%	Total: 0.84 CAC ≥ 400: 0.71; CAC 0-400: 0.85
De Geer et al[31], 2019	American Journal of Roentgenology	No	The MACHINE registry	cFFR, version 2.1, Siemens	351	525	No	Total: 78%; 80 kv: 86%; 100 kv: 77%; 120 kv: 78%	Total: 0.84; 80 kv: 0.90; 100 kv: 0.82; 120 kv: 0.84
Xu et al[33], 2020	European Radiology	No	10 individual centers across China	cFFR, version 3.2.0, Siemens	437	570	No	Total: 89%; High quality: 94%; Low quality: 83%	Total: 0.89; High quality: 0.93; Low quality: 0.80
Kumamaru et al[28], 2020	European Heart Journal - Cardiovascular Imaging	No	Multi-center	Python 3.6	131	-	No	Per-patient: 76%	Per-patient: 0.78
Li et al[29], 2021	Acta Radiologica	No	Single center	DEEPVESSE-FFR Platform	73	85	No	Per-lesion: 92%; Per-patient: 91%	Per-lesion: 0.96
Xu et al[35], 2020	European Radiology	No	A Chinese multicenter study	cFFR, version 3.1.0, Siemens	442	544	No	Per lesion: 90%	-

IRIS: Iterative reconstruction in image space; FBP: Filtered back projection; CAC: Coronary artery calcium; FFR: Fractional flow reserve; AUC: Area under the curve; CT: Computed tomography.

In addition, the influences of CT reconstruction algorithms, image quality, tube voltage, coronary calcium, and gender on the diagnostic performance of FFR-CT_ML were investigated in several studies. In a sub-study of MACHINE Registry, Tesche et al[30] examined the impact of calcification on CT-FFR_ML determination and concluded that CT-FFR_ML revealed a statistically significant different (P = 0.04) performance as Agatston calcium score increased: The AUC in high Agatston scores (CAC ≥ 400) was 0.71 (95%CI: 0.57-0.85) and in low-to-intermediate Agatston scores (CAC > 0 to < 400) was 0.85 (95%CI: 0.82-0.89). In another sub-study of MACHINE Registry, De Geer et al[31] examined the impact of different tube voltages on CT-FFRML determination and concluded that performance does not vary significantly between tube voltages of 100 kVp (AUC: 0.82) and 120 kVp (AUC: 0.84), while the AUC was 0.90 in examination with a tube voltage of 80 kVp. Based on data of the MACHINE Registry, Baumann et al[32] evaluated the impact of gender on the performance of FFRCTML and they found that FFRCTML performs equally in men and women (both with an AUC of 0.83). In a retrospective Chinese multicenter study, Xu et al[33] investigated the effect of image quality on the diagnostic performance of FFRCTML in 437 patients with 570 vessels. They found that the AUC of high-quality images [0.93 (95%CI: 0.88-0.98), n = 159] was significantly (P = 0.02) superior to that of low-quality images [0.80 (95%CI: 0.70-0.90), n = 92]. And CCTA with a score ≥ 3, intracoronary enhancement degree of 300–400 HU, and heart rate below 70 bpm at scanning could be of great benefit to more accurate FFRCTML analysis. In a retrospective single center study, Mastrodicasa et al[34] evaluated the influence of different CT reconstruction algorithms on the performance of CT-FFR_ML in 40 CCTA datasets. CT-FFRML values were significantly different between iterative reconstruction in image space (IRIS) and filtered back projection algorithms, whereas no difference was observed in diagnostic accuracy (both 81.8%, P = 1.000). Additionally, they found that IRIS improved CT-FFRML post-processing speed significantly.

It should be mentioned that CT-FFR_ML value for each location along the coronary is trained when taking the CT-FFR_CFD as ground truth. Although the diagnostic accuracy of CT-FFR derived using deep learning (DL) methods was validated in several studies, it is still susceptible to the CCTA scanning factors. In the future, more attention should be paid to the widespread use of a local software solution that allows for image-variation and user-variation.

OTHER AI ALGORITHMS FOR PREDICTION OF MYOCARDIAL ISCHEMIA

Except for the ML based FFR-CT platforms described above, some other AI algorithms were developed recently for prediction of myocardial ischemia. These approaches are in early stage but show better interpretability, which were established via an integration of qualitative or quantitative features derived from CCTA images and clinical factors.

In 2018, Dey et al[37] developed an integrated ML ischemia risk score (ML-IRS) from quantitative plaque measures using a supervised learning process to predict functionally significant stenosis in a prospective multicenter trial of 254 patients with 484 vessels. The ML-IRS exhibited a higher AUC (0.84) than conventional CCTA measures, including stenosis (0.76), LD-NCP volume (0.77), total plaque volume (0.74), and pre-test likelihood of CAD (0.63), for predicting lesion-specific ischemia by invasive FFR. Thereafter, the ML-IRS was integrated into coronary plaque analysis research software for generating a percent probability of pathological FFR on CCTA data.

In 2019, van Hamersvelt et al[38] proposed a DL method based on the LVM in resting CCTA images to identify functionally significant coronary artery stenosis using 126 patients. The DL approach achieved a higher AUC of 0.76 compared to degree of stenosis (AUC = 0.68).

In 2020, Shu et al[39] established a radiomics nomogram based on myocardial segments for predicting chronic myocardial ischemia using multivariate logistic regression. The accuracy of the nomogram for distinguishing chronic myocardial ischemia from normal myocardium was 0.839, 0.832, and 0.816 in the training, test, and validation cohorts, respectively.

PROGNOSTIC SIGNIFICANCE

PVAT-based radiomics for improving cardiac risk prediction

Early detection of vascular inflammation, which is a major contributor to atherogenesis and atherosclerotic plaque rupture[40,41], would enable better cardiovascular risk stratification[42]. The vascular inflammation can be detected by characterizing the phenotypic changes in PVAT using the fat attenuation index (FAI) in routine CCTA images[43,44]. FAI was defined as the average attenuation of all voxels with attenuation values between -190 HU and -30 HU located within a radial distance from the outer coronary artery wall equal to the average diameter of the respective vessel, as described previously[43,44]. However, FAI is an average of the voxel intensity values and does not account for the complex spatial relationship among voxels.

Recently, some studies investigated whether radiomics analysis could help to extract more information from the PVAT that cannot be captured by human eyes. The radiomics features surrounding PVAT mainly include two parts: (1) PVAT surrounding the standardized coronary segments, which was often investigated at a per-patient level; and (2) PVAT around the target lesion, which was at a per-lesion level.

As for the per-patient level, Oikonomou et al[45] developed an AI-powered radiotranscriptomic signature for predicting cardiac risk based on the radiomics features extracted from PVAT around the proximal to distal right coronary artery (RCA) and the left coronary artery in CCTA images. A fat radiomic profile (FRP) was established, using random forest model based on the features extracted from the standardized coronary segments, to distinguish the 101 patients who experienced major adverse cardiac events (MACE) within 5 years from 101 matched controls. The FRP was significantly associated with the risk of MACE [adjusted hazard ratio (HR): 1.12, 95%CI: 1.08-1.15, P < 0.001]. And patients with an FRP ≥ 0.63 had a 10.8-fold higher risk of MACE than those with an FRP < 0.63, after adjusted for clinical factors. The AUC of FRP in predicting MACE was 0.774 (95%CI: 0.622-0.926) in the external validation dataset (20% of the 202 samples). When added to the traditional model, FRP improved the distinguishing performance from an AUC of 0.754 to 0.880. Additionally, they found that FRP was significantly higher in 44 patients with acute MI compared with 44 controls (P < 0.001), but unlike FAI, FRP remained unchanged 6 mo later in 16 patients with acute MI (AMI), confirming that FRP detects persistent PVAT changes that cannot be captured by FAI.

As for the per-lesion level, in 2020, Lin et al[46] further explored the prognostic value of the radiomics features of PVAT around not only the standardized coronary segments but also lesions in a prospective case-control study. They found no significant difference between the PVAT radiomics features of culprit and non-culprit lesions in patients with AMI, lending further support to the pan-coronary inflammatory hypothesis. But on the other hand, as for the per-patient level, patients with AMI (n = 60) have a distinct PVAT radiomics phenotype surrounding the proximal RCA compared with patients with stable (matched, n = 60) or no CAD (matched, n = 60). Among the three models that they developed, the PVAT-based radiomics model (AUC: 0.87) outperforms the clinical model (AUC: 0.76) and the combined model incorporating clinical factors and PVAT attenuation (AUC: 0.77) in identifying AMI with stable CAD and controls. Additionally, after a 6-mo follow-up of patients with AMI, no significant change was observed in the radiomics features of PVAT surrounding the proximal RCA or non-culprit lesions.

QUANTITATIVE CT FEATURES-BASED ML FOR OUTCOME PREDICTION

Information extracted from CCTA images along with other clinical factors are associated with prognosis, and AI technology demonstrated great potential to enhance decision-making and improve patient outcomes. Currently, the prognostic value of ML algorithms using quantitative CCTA features together with clinical variables was investigated by researchers in several studies[47-53], in which promising results were obtained. The ML algorithms performed better than traditional predictors, not only for short-term treatment decisions but also for long-term risk predictions, as summarized in Table 2.

Table 2 summary of the current literature on the prognostic value of machine learning algorithms in coronary computed tomography angiography.

Ref.	Journal	Prospective	Multi Center	No. of Patients	No. of Events	Algorithm	Endpoint	Follow-up time	Performance
Motwani et al[48], 2017	European Heart Journal	Yes	Yes	10030	745 died	LogitBoost	5-yr all-cause mortality	5.4 ± 1.4 yr	AUC = 0.79
van Rosendael et al[47], 2018	Journal of Cardiovascular Computed Tomograph	Yes	Yes	8844	350 death and 259 non-fatal MI	XGBoost	MI and death	4.6 ± 1.5 yr	AUC = 0.77
Johnson et al[49], 2019	Radiology	No	No	6892	380 died of all causes and 70 died of CAD	Logistic regression, KNN, Bagged trees, and classification neural network	Death or cardiovascular events	9.0 yr (interquartile range, 8.2–9.8 yr)	For all-cause mortality: AUC = 0.77; For CAD deaths: AUC = 0.85
van Assen et al[50], 2019	European Journal of Radiology	No	No	45	16 MACEs	Regression analysis	MACE	12 mo	AUC = 0.94
von Knebel Doeberitz et al[51], 2019	The American Journal of Cardiology	No	No	82	18 MACEs	Integration of CT-FFR, stenosis ≥ 50% and plaque markers	MACE	18.5 mo (interquartile range 11.5 to 26.6 mo)	AUC = 0.94
Commandeur et al[52], 2020	Cardiovascular Research	Yes		1912	76 MI and/or cardiac death	ML	Long-term risk of MI and cardiac death	14.5 ± 2 yr	AUC = 0.82
Kwan et al[53], 2021	European Radiology	Yes	Yes	352		ML	Future revascularization		AUC = 0.78

XGBoost: Extreme gradient boosting; KNN: K-nearest neighbors; ML: Machine learning; AUC: Area under the curve; MACE: Major adverse cardiac events; CT-FFR: Computed tomography-fractional flow reserve; CAD: Coronary artery disease; MI: Myocardial infarction.

One of the first major studies using CCTA based ML approach for prognosis evaluation is a large prospective multi-center study conducted by Motwani et al[48] in 2017. They developed an ML model in CCTA to predict 5-year all-cause mortality using a dataset of 10030 patients with suspected CAD from the CONFIRM registry (Coronary CT Angiography Evaluation for Clinical Outcomes: An International Multicenter). The ML model was established after an automated feature selection procedure based on 44 CCTA-derived parameters and 25 clinical parameters. One summary score for clinical parameters (Framingham risk score, FRS) and three composite CCTA-based scores [including the segment stenosis score (SSS), the segment involvement score (SIS), and the modified Duke prognostic CAD index (DI)] were derived. The ML model exhibited a significant higher AUC compared with the conventional scores alone for predicting 5-year all-cause mortality (ML: 0.79 vs FRS: 0.61, SSS: 0.64, SIS: 0.64, and DI: 0.62; P < 0.001).

Two years later, in 2019, Johnson et al[49] developed another ML model using 64 vessel features derived from CCTA images, to discriminate between patients with and without subsequent death or cardiovascular events in a retrospective single-center study with 6892 patients. The performance of the ML model was compared with that of Coronary Artery Disease Reporting and Data System (CAD-RADS) score. For prediction of all-cause mortality, the AUC of the ML model was significantly higher than that of CAD-RADS (0.77 vs 0.72, P < 0.001). For prediction of coronary artery deaths, the AUC was significantly higher for the ML model than for CAD-RADS (0.85 vs 0.79, P < 0.001).

In 2020, Commandeur et al[52] developed an ML model integrating clinical parameters with quantitative imaging-based variables for predicting events of long-term risk of MI and cardiac death in asymptomatic subjects using the dataset with 1912 cases from the randomized EISNER trial. The ML model obtained a significantly higher AUC than atherosclerotic cardiovascular disease (ASCVD) risk and CAC score for predicting events (ML: 0.82; ASCVD: 0.77; CAC: 0.77; P < 0.05). Subjects with a higher ML score had a significant high hazard of suffering events (HR: 10.38, P < 0.001).

As for the short-term decision-making, in 2020, Kwan et al[53] examined whether the ML-IRS, developed by Dey et al[37] in 2018, as described previously (Figures 1 and 2), can predict revascularization in patients referred to ICA after CCTA in a prospective dual-center study of 352 patients with 1056 analyzable vessels. It would be beneficial to effectively identify the patients who were referred for standard clinical CCTA followed by ICA due to decision by a primary treating physician but did not receive revascularization, because those patients are a high-cost population with low yield from the invasive procedure. The results indicated that ML-IRS, when added to the traditional risk model, significantly improve the prediction of future revascularization with an increased AUC from 0.69 (95%CI: 0.65-0.72) to 0.78 (95%CI: 0.75-0.81) (P < 0.0001).

Overall, the application of AI in CCTA has a potential future for improving the short-term risk stratification and long-term prognostic evaluation. The ML algorithms that have been proposed should be validated and tested in real world with larger external cohorts including diversity of patients so as to make sure the models be optimized and generalized.

CONCLUSION

Current AI applications in CCTA images are mostly designed in two dimensions: (1) For the radiologists, AI is applied to improve efficiency and reduce workload via optimizing the clinical workflow, such as improvement of image reconstruction from lower quality to high quality (e.g., low-dose acquisition or motion artifacts) and structured reporting; and (2) For the patients, AI is utilized to increase benefit and improve prognostic evaluation via providing valuable diagnostic information more accurately, such as detection of anatomic and functional stenosis, quantification of plaques, and estimation of the vascular inflammation.

In this review, we mainly focused on the second dimension which is patient oriented. AI algorithms in CCTA images provide information in a more objective, reproducible, and rational manner compared to human perception, and exhibits its potential to outperform human in several cardiac fields. However, CCTA imaging lagged behind cancer imaging in the clinical translational of AI-based methods, especially the radiomics analysis. It has long been demonstrated in the field of cancer imaging that radiomics signatures are superior to traditional factors in predicting outcomes of patients. But only a few studies using radiomics analysis have been conducted in CCTA images. Considering that regions of interest (ROIs) segmented before the extraction of radiomics features, can be drawn along the edge of the tumor in cancer imaging generally, in CCTA images the selection of ROIs brings about challenges. Researchers hereby performed radiomics analysis around the PVAT or LVM or plaques. And recently, several groups succeeded in developing automated segmentation of PVAT and LVM, which provides probabilities to explore more novel non-invasive predictors for improvement of risk stratification and prognosis in patients with CAD.

Additionally, FFR-CT driven by AI is a hot topic in recent years. Various FFR-CT platforms are developed and adding into the clinical diagnostic workflow for not only research purpose but also commercial use. In the near future, the FFR-CT platforms would bring major changes in the way to make decisions for patients with CAD before invasive coronary angiography.

However, before AI solutions can be truly widely implemented in daily clinical workflow or the reading room, several issues should be noted: (1) The algorithms need to be carefully validated in multi-center studies or large clinical trials to ensure the robustness and generalization; (2) The approval of clinical application is required to prove the accuracy and safety of the AI products; and (3) The legal and ethical issues should be taken into consideration.

In summary, AI offers the possibility to optimize clinical workflow and provide precise information for diagnostic and treatment, which will benefit both radiologists and patients. However, it is pertinent to note that AI will not simply substitute the cardiac radiologists, and human support or supervision is still needed. Rather, the cardiac radiologists need to be fully aware of the strengths and limitations of AI.

ACKNOWLEDGEMENTS

First, I want to show my great gratitude to my teacher Dr. Hou, a responsible and respectable scholar, offering valuable suggestions for revision. In addition, I am grateful to Guo Y for her contribution to the writing process. Also, I wish to thank those who have offered me great help and support, such as editors, reviewers, and publishers.