Table 3.
Main AI applications to stress cardiac magnetic resonance (S-CMR).
| References | Summary | Performance |
|---|---|---|
| Assessment of cardiac function | ||
| Bai et al. (35) | Automated myocardial segmentation using a DL algorithm trained in a huge dataset (>4,500 subjects) | Excellent correlation with manual measurement (Dice's coefficient 0.94 for the LV cavity, 0.88 for the LV myocardium and 0.90 for the RV cavity) |
| Curiale et al. (36) | Automated LV quantification using DL | Good accuracy for myocardial segmentation (Dice's coefficient 0.9); high correlation index for LVEDV and LVESV (0.99), LV EF (0.95), and for SV and CO (0.93). |
| Tissue characterization | ||
| Kotu et al. (37) | Arrhythmic risk stratification of CAD patients through the radiomic analysis of the scar tissue | Highly accurate (94%) classification of CAD patients in high- and low arrhythmic risk groups |
| Xu et al. (38) | Automated detection of MI | 94% overall accuracy in detecting the MI area extension, position and shape |
| Larroza et al. (39) | Distinction of acute and chronic MI on CMR-LGE and non-enhanced CMR through an ML model combined with radiomics | High AUC, sensitivity and specificity in the distinction between acute and chronic MI both on CMR-LGE (0.86, 0.81, and 0.84, respectively) and on non-enhanced CMR (0.82, 0.79, and 0.80, respectively) |
| Larroza et al. (40) | Automated identification of myocardial transmural scar on non-enhanced CMR | Sensitivity of 92% for transmural scar identification |
| Baessler et al. (41) | Automated scar detection on non-enhanced CMR images with a combined ML and radiomics algorithm | Identification of five independent texture features, which allowed scar identification. The best features combination allowed an AUC of 0.93 and 0.92 for diagnosing large and small MI, respectively |
| Moccia et al. (42) | Comparison of two DL scar segmentation protocols for automated scar detection on CMR-LGE images | 88% median sensitivity and 71% DICE similarity coefficient by the protocol that limited the analysis to the myocardial region. |
| Zabihollahy et al. (43) | Semiautomated DL method for LV myocardial scar segmentation from 3D CMR-LGE images. | 94% DICE similarity coefficient for LV myocardial scar segmentation |
| Zhang et al. (44) | Automated detection, localization and quantification of myocardial fibrosis on non-enhanced CMR | No difference between non-enhanced cardiac cine and CMR-LGE analyses: number of scar segments (p = 0.38), mean per-patient scar area (p = 0.27) percentage of damaged myocardial tissue (p = 0.17) |
| Ma et al. (45) | Combination of radiomics and T1 mapping for the automated identification of MVO | Radiomics combined with T1 values compared to T1 values alone better identified MVO (AUC 0.86) and showed higher predictive value for LV longitudinal systolic myocardial contractility recovery (AUC 0.77). |
| Perfusion S-CMR | ||
| Scannell et al. (46) | Automated processing and segmentation myocardial perfusion data on S-CMR | High accuracy compared to manual processing and segmentation (Dice similarity coefficient for myocardial segmentation 0.8) |
| Xue et al. (47) | Automated assessment of MBF on S-CMR | High accuracy compared to manual analysis in myocardial segmentation (Dice similarity coefficient 0.93). No difference in the per-sector MBF identification (p = 0.92) |
CAD, coronary artery disease; CO, cardiac output; EF, ejection fraction; LGE, late-gadolinium enhancement; LV, left ventricle; LVEDV, left ventricle end diastolic volume; LVESV, left ventricle end systolic volume; MBF, myocardial blood flow; MI, myocardial infarction; MVO, microvascular obstruction; SV, stroke volume; SVM, support vector machine.