Development of artificial intelligence in epicardial and pericoronary adipose tissue imaging: a systematic review

Background Artificial intelligence (AI) technology has been increasingly developed and studied in cardiac imaging. This systematic review summarizes the latest progress of image segmentation, quantification, and the clinical application of AI in evaluating cardiac adipose tissue. Methods We exhaustively searched PubMed and the Web of Science for publications prior to 30 April 2021. The search included eligible studies that used AI for image analysis of epicardial adipose tissue (EAT) or pericoronary adipose tissue (PCAT). The risk of bias and concerns regarding applicability were assessed with the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool. Results Of the 140 initially identified citation records, 19 high-quality studies were eligible for this systematic review, including 15 (79%) on the image segmentation and quantification of EAT or PCAT and 4 (21%) on the clinical application of EAT or PCAT in cardiovascular diseases. All 19 included studies were rated as low risk of bias in terms of flow and timing, reference standards, and the index test and as having low concern of applicability in terms of reference standards and patient selection, but 16 (84%) studies did not conduct external validation. Conclusion AI technology can provide accurate and quicker methods to segment and quantify EAT and PCAT images and shows potential value in the diagnosis and risk prediction of cardiovascular diseases. AI is expected to expand the value of cardiac adipose tissue imaging. Supplementary Information The online version contains supplementary material available at 10.1186/s41824-021-00107-0.

tissue. The reference lists of relevant studies and reviews were checked to find other possible studies. The language was restricted to English. The detailed search strategies are listed in the supplement.

Selection process and data collection
After removing duplicate articles, two reviewers (L.Z. and BB.J.) independently reviewed the titles, abstracts, and full texts to select eligible articles. Articles that did not use AI or that did not involve EAT or PCAT were excluded. The same two reviewers independently extracted the data from the included articles, including the article characteristics (authors, year of publication, and research purposes), sample sizes, imaging equipment, AI methods, and main results. In the process, the two reviewers resolved their disagreement via consensus with the third senior reviewer (XQ.X.).

Study risk of bias assessment
Two reviewers (L.Z. and BB.J.) independently assessed the risk of bias and concerns regarding applicability of the included studies using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool (Whiting et al. 2011) and resolved the disagreement by consultation with the third senior reviewer (XQ.X.). QUADAS-2 consists of four key domains: (1) patient selection, (2) index test, (3) reference standard, and (4) flow and timing. Considering the concerns regarding applicability in the index test domain, the studies with external validation were considered low concern.

Study selection and study characteristics
The initial search identified 140 citation records. Of these, 45 duplicate records were deleted, and 56 records were removed by browsing the titles and abstracts. After reading the full texts, 22 studies were excluded. Thereafter, 19 studies were finally included in the systematic review. Figure 1 depicts the flowchart of the selection process. Figure 2 and Table 1 show the characteristics of the included studies. Of the 19 included studies, 15 (79%) used AI technology to segment and quantify EAT or PCAT, while 4 (21%) explored the clinical value of AI in the diagnosis and risk prediction of cardiovascular diseases. In terms of imaging equipment, 18 studies (95%) used computed tomography (CT), including 12 (63%) applying non-contrast scans and 6 (32%) performing coronary CT angiography (CCTA) scans. Only one study (5%) used cardiac magnetic resonance (CMR) imaging.  Table S1 represent the results of risk of bias and concerns regarding applicability according to the QUADAS-2 tool. For the assessment of bias risk, all 19 studies were rated as low risk in terms of flow and timing, reference standards, and index tests. In terms of patient selection, 5 studies (26%) were rated as low risk while 14 (74%) were rated as unclear risk. For the assessment of the concerns of applicability, all 19 studies were rated as low concern in terms of reference standards and patient selection. However, for the index test, only one study (5%) was rated as low concern because of the existence of external validation; 16 studies (84%) were considered unclear because they only conducted internal validation (cross or leave-one-out validation), and 2 studies (11%) were rated as high concern due to a lack of validation.

Discussion
This systematic review summarized the research progress of AI in EAT and PCAT image analysis. An exhaustive literature search indicated that AI can improve the image segmentation and quantification of EAT or PCAT and can also be used as a tool for   the diagnosis and risk prediction of coronary artery diseases based on EAT or PCAT. However, most of the studies have not applied external validation, which suggests that the generalizability of these studies remains to be improved.

Segmentation of EAT and PCAT
Given the potential clinical value of EAT, the segmentation of EAT images is the crucial step for further quantitative analysis. In the traditional analysis mode, radiologists have to delineate the boundary of EAT manually. This procedure is user-dependent, time-consuming, and poorly reproducible, so it is necessary to develop an accurate, quicker, and reproducible method for EAT segmentation (Norlén et al. 2016;Commandeur et al. 2018). Researchers have developed machine learning algorithms to segment EAT. Rodrigues et al. utilized a genetic algorithm to optimize the parameters of an ellipse that was used to simulate the pericardium contour (Rodrigues et al. 2017b). After 10, 100, and 200 generations of genetic algorithm iteration, the percentage of epicardial fat engulfed by the ellipse was 97.3%, 98.8%, and 99.5%, respectively. Norlén et al. proposed a method to detect the pericardium using the random forest classification algorithm (Norlén et al. 2016). This method used feature-based multi-atlas regulations for spatial initiation in CCTA images. Finally, the  segmentation was completed in global optimization through graph cuts, and the Dice similarity coefficient (DSC) between the proposed method and experts was 0.91. The average segmentation time was 51.9 s. The DSC is a commonly used measurement to evaluate the similarity between two structures, and its value ranges from 0 to 1. A higher DSC represents a higher similarity of image segmentation (Andrews and Hamarneh 2015).
Researchers have also explored the utilization of deep learning in EAT segmentation. Zhang et al. developed a dual U-Net CNN for the automatic segmentation and quantification of EAT (Zhang et al. 2020). Compared with a single U-Net (DSC = 0.766) and Seg-Net (DSC = 0.767), this method segmented the EAT of 20 patients with a mean DSC of 0.912.
In addition to tracking PCAT manually on axial images, PCAT can also be automatically segmented in 3D space. PCAT is considered adipose tissue whose radial distance from the outer wall of the coronary artery is equal to the mean diameter of the artery. A welldeveloped software package (Aquarius Workstation, TeraRecon GmbH) automatically segments PCAT radially outwards from the outer wall of coronary vessels in 3D space (Antonopoulos et al. 2017). PCAT is then defined by the CT value range (typically from − 190 HU to − 30 HU) of the adipose tissue within the segmented volume of interest.
In short, researchers have developed several AI-based methods for the automatic segmentation of EAT and PCAT. The studies applying AI technology to evaluate EAT (Hasebe et al. 2020) and PCAT (Oikonomou et al. 2019) showed good interobserver agreement, and the correlation coefficients were 0.93 and 0.94, respectively. These AIbased methods have the advantages of high accuracy, short processing time, and good reproducibility, which lay the foundation for further research.

Quantification of EAT and PCAT
Quantification follows image segmentation. The quantitative parameters for EAT or PCAT are generally volume, area, and thickness. Commandeur et al. proposed a new multi-task framework using deep CNNs to quantify epicardial and thoracic adipose tissues . This method showed that the automatic volume quantification of epicardial and thoracic adipose tissue was strongly consistent with the manual method, and the median DSC was 0.823 and 0.905, respectively. The correlation coefficient between automatic and manual EAT volume quantification was 0.924. The automatic method for quantifying EAT took 25.63 ± 3.72 s per case, much shorter than the 10-11 min of the manual method. Figures 4 and 5 represent the image segmentation and quantification of EAT and PCAT, respectively.

Clinical applications of machine learning in EAT
EAT plays an important role in the development of CVD through pathophysiological mechanisms (Mazurek et al. 2003;Cheng et al. 2008). Evidence suggests that abundant EAT is a predictor of adverse cardiovascular events (Mahabadi et al. 2013;Goeller et al. 2018). Since machine learning algorithms can maximize the data mining of image features and clinical factors, researchers have used them to analyze EAT. In a prospective study, Commandeur et al. developed a machine learning-derived risk score based on the extreme gradient boosting (XGBoost) algorithm, which combined clinical risk factors with the coronary artery calcium (CAC) score and automatically quantified EAT to predict the long-term risk of myocardial infarction (MI) and cardiac death Zhang et al. European Journal of Hybrid Imaging (2021) (Commandeur et al. 2020). The machine learning-derived risk score (AUC = 0.82) performed better than the risk factors for atherosclerotic CVD (AUC = 0.77) and CAC score (AUC = 0.77). Otaki et al. applied the same algorithm (XGBoost) and found that the epicardial fat volume was an independent predictor of impaired myocardial flow reserve (Otaki et al. 2015). Compared with the CAC score (AUC = 0.66) and epicardial fat volume (AUC = 0.67), the AUC of the combined risk score integrating the CAC score and epicardial fat volume was 0.73, which improved the prediction ability of impaired myocardial flow reserve. In general, the existing evidence suggests that EAT, with the support of machine learning, can improve the diagnosis and prognostic prediction of CVD.

Clinical applications of machine learning in PCAT
PCAT is the adipose tissue enveloping the coronary artery, which involves the evolution of atherosclerosis (Dey et al. 2010). The pericoronary fat attenuation index (FAI) is a novel imaging biomarker that maps the spatial changes in PCAT attenuation in CCTA images. The FAI has been used to identify subclinical coronary artery disease (Antonopoulos et al. 2017) and improve the risk prediction of all-cause cardiac mortality (Oikonomou et al. 2018). In addition to the FAI that simply represents the CT value, radiomics can mine the high-throughput quantitative information of the PCAT. Oikonomou et al. reported that the radiomics signature of the PCAT was associated with the fibrosis and vascularity of the PCAT, and the PCAT represented the adipose tissue remodelling caused by coronary inflammation (Oikonomou et al. 2019). In addition, they proposed a random forest method to analyze 1391 radiomic features of the PCAT extracted from 101 patients who had major adverse cardiac events (MACEs) within 5 years after undergoing CCTA and 101 matched controls. Their machine learning-based model performed better than the existing clinical risk factor model (Δ[C-statistic] = 0.126, p < 0.001) in predicting the outcome of MACEs. Lin et al. explored the ability of the radiomics signature of the PCAT in CCTA images to discriminate patients with MI from those with stable or no coronary artery disease (Lin et al. 2020). Using an XGBoost that combined clinical factors, PCAT attenuation, and radiomics features, their method significantly improved the discrimination ability of acute MI (AUC = 0.87) compared with a model with clinical factors and PCAT attenuation (AUC = 0.77) or clinical factors alone (AUC = 0.76). In brief, the implementation of machine learning can maximize the information extracted from PCAT and play an important role in the diagnosis and risk prediction of coronary artery disease.

Multi-modal imaging of cardiac adipose tissue
Multiple types of equipment can be used to evaluate EAT, including CT, CMR, and echocardiography. CT offers volumetric visualization and quantification of cardiac adipose tissue with high spatial resolution. However, CT has the disadvantage of unavoidable radiation exposure. CMR is considered to be the best imaging method for adipose tissue because of its superb fat signal display (Wong et al. 2017). According to the results of Mahajan et al., there was a significant correlation between CMR and autopsy for measuring the mass of adipose tissue with an intraclass correlation coefficient > 0.8 (Mahajan et al. 2013). The main disadvantages of CMR are its limited availability and high costs. Echocardiography is widely used, but the image quality is poor and not conducive to quantitative analysis. The quantification of EAT by echocardiography was limited to the thickness of adipose tissue on the right ventricular free wall (related to EAT thickness on CMR: R = 0.905) (Iacobellis et al. 2003;Gaborit et al. 2017). Compared with the CT widely used in EAT analysis, research based on echocardiography and CMR is limited. In the included studies, only one used deep learning to segment EAT in CMR images with a low DSC of 0.56. More studies are expected to investigate EAT using AI and multi-modal imaging.

Future prospects
The segmentation and quantification of EAT is an important basis of EAT research. Researchers have developed AI-based methods to make this process more accurate, quicker, and reproducible. In addition to developments in the experimental environment, more robust and generalizable AI methods are needed to assist the clinical application of EAT analysis.
As a booming field in recent years, radiomics enables the in-depth mining of image information. Based on EAT and PCAT, scholars have begun to use radiomics to diagnose and predict the risk of cardiovascular disease. However, to validate the clinical value of radiomics-based imaging biomarkers, more well-conducted multi-center studies are needed. Deep learning has become an auxiliary tool for image segmentation. On the basis of automatic segmentation of PCAT, an AI algorithm was used to quantitatively analyze and model the FAI, which expands the research of cardiac imaging. In addition to image segmentation, more AI-based research is expected to be launched to further explore the diagnostic and prognostic value of EAT and PCAT in CVD.

Conclusion
With the recent development of AI, researchers have begun to apply AI to analyze EAT and PCAT images. AI-based methods provide accurate and quicker image segmentation and quantification. Although the publications are not numerous, the published studies have suggested the potential clinical application of EAT and PCAT in the diagnosis and risk prediction of CVD. Although AI research on cardiac adipose tissue is still in its infancy, the application of AI technology is expected to expand the clinical value of cardiac adipose tissue in cardiac imaging.
Additional file 1: Table S1. Assessment of risk of bias and concerns regarding applicability according to the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool.