Table 1 Characteristics of studies included in the systematic review
Authors | Research purpose | Patients, n | Scan mode | AI method | Performance |
|---|---|---|---|---|---|
(Rodrigues et al. 2015) | Automated segmentation of epicardial and mediastinal fats | 20 | Non-contrast CT | ML: Intersubject registration + RF | DSC = 0.968 |
(Rodrigues et al. 2016) | Automatic segmentation and quantification of cardiac fats | 20 | Non-contrast CT | ML: Atlas-based + RF | DSC = 0.977 |
(Rodrigues et al. 2017b) | Automated segmentation of epicardial fat | 20 | Non-contrast CT | ML: Genetic algorithms | The percentage of epicardial fat engulfed by the ellipse was 99.5% |
(Norlén et al. 2016) | Automatic segmentation and quantification | 30 | CCTA | ML: Multi-atlas + RF + Markov random field | CC = 0.99 DSC = 0.91 |
(Zlokolica et al. 2017) | Semiautomatic EAT segmentation | 10 | CCTA | ML: Fuzzy c-means clustering + geometric ellipse fitting | DSC = 0.69 |
(Commandeur et al. 2018) | Segmentation and quantification of EAT | 250 | Non-contrast CT | DL: CNN | CC = 0.924 DSC = 0.823 |
(Commandeur et al. 2019) | Quantification of EAT | 776 | Non-contrast CT | DL: CNN | DSC = 0.871 |
(Li et al. 2019) | Automatic pericardium segmentation | 53 | Non-contrast CT | DL: U-Net | AUC = 0.87 |
(Aarthy et al. 2019) | Quantification of EAT | 20 | Non-contrast CT | DL: K mean clustering + CNN | CC = 0.803 |
(Fulton et al. 2020) | Segmentation of EAT | 32 | Cardiac magnetic resonance imaging | DL: Neural network | DSC = 0.56 ± 0.12 |
(Zhang et al. 2020) | Automatic epicardial fat segmentation and quantification | 20 | Non-contrast CT | DL: dual U-Nets + morphological processing layer | CC = 0.93 DSC = 0.91 |
(He et al. 2020a) | Automatic segmentation and quantification of EAT | 200 | CCTA | DL: 3D deep attention U-Net | DSC = 0.927 |
(He et al. 2020b) | Automatic quantification of myocardium and pericardial fat | 422 | CCTA | DL: Deep attention U-Net | ICC = 0.97 DSC = 0.88 |
(Otaki et al. 2015) | Prediction of impaired myocardial blood flow from clinical and imaging data (EFV) | 85 | Non-contrast CT | ML: Ensemble-boosting logitboost algorithms | AUC = 0.73 vs 0.67 (ML vs EFV) |
(Rodrigues et al. 2017a) | Prediction of epicardial and mediastinal fat | 20 | Non-contrast CT | ML: Rotation forest + multi-layer perception regressor | Predicting mediastinal fat based on EAT: CC = 0.986 RAE = 14.4% Predicting EAT based on mediastinal fat: CC = 0.928 RAE = 32.5% |
(Commandeur et al. 2020) | Predict the long-term risk of MI and cardiac death based on clinical risk, CAC, and EAT | 1912 | Non-contrast CT | ML: XGBoost | ML-AUC = 0.82 CAC-ACU = 0.77 ASCVD-AUC = 0.77 |
(Tamarappoo et al. 2021) | The long-term prediction of hard cardiac events | 1069 | Non-contrast CT | ML: XGBoost | ML-AUC = 0.81 CAC-AUC = 0.81 ASCVD-AUC = 0.74 |
(Oikonomou et al. 2019) | Radiotranscriptomic signature of perivascular fat improves cardiac risk prediction | 1575 | CCTA | ML: Radiomics-RF | For MACE discrimination: with radiomics signature-AUC = 0.88 without radiomics signature-AUC = 0.754 |
(Lin et al. 2020) | Radiomics analysis of PCAT to distinguish patients with MI | 177 | CCTA | ML: XGBoost | ML-AUC = 0.87 clinical features + PCAT attenuation-AUC = 0.77 clinical features alone-AUC = 0.76 |