Table 5.
Main AI applications to Coronary Computed Tomography Angiography (CCTA).
| References | Summary | Performance |
|---|---|---|
| Coronary stenoses grading | ||
| Kelm et al. (64) | Automated ML detection, grading and stenoses grading on CCTA images | Good sensitivity and specificity (95 and 67%) compared to expert evaluation to correctly detect significant coronary artery stenoses |
| Kang et al. (65) | Automated ML detection of coronary artery stenoses on CCTA images | High sensitivity (93%), specificity (95%), and accuracy (94%), with AUC (0.94) for coronary artery stenoses detection compared to experts' visual assessment |
| Zreik et al. (66) | Automated LV myocardium analysis to identify patients with significant coronary artery stenoses | The DL application correctly performed LV segmentation (Dice similarity coefficient 0.91) and identified patients with significant coronary artery stenosis with an AUC value of 0.74 |
| Hong et al. (67) | Automated DL coronary artery stenoses grading (plaque segmentation, MLA and percent DS quantification) on CCTA images | Excellent correlation of ML performance to expert readers (ρ = 0.984 for MLA; ρ = 0.957 for DS p < 0.001 for all) |
| Muscogiuri et al. (68) | Automated DL classification of coronary artery stenoses according to CAD-RADS | The DL algorithm showed its best performance in differentiating between CADRADS 0 (i.e., no coronary atherosclerosis) vs. CADRADS > 0 (i.e., detectable coronary atherosclerosis) with a sensitivity of 66% and a specificity of 91%, compared to experts' analysis |
| Plaque phenotype characterization | ||
| Dey et al. (69) | Automated distinction between calcified and non-calcified plaques | Strong correlation between automated plaque analysis and expert readers (ρ = 0.94, for NCP volume; ρ = 0.88, for CP volume; ρ = 0.90 for NCP and CP composition) |
| Kolossváry et al. (70) | Identification of radiomic features associated to the presence of NRS in coronary artery plaques | Identification of NRS through radiomic analysis with an AUC > 0.92. One radiomic feature reached a remarkable AUC of 0.92 for NRS identification |
| Masuda et al. (71) | Automated ML algorithm for the detection of fibrous or fibro-fatty coronary artery plaques | The ML algorithm identified high risk coronary plaques better than intravascular ultrasound evaluation (AUC 0.92 vs. 0.83) |
| Zreik et al. (72) | DL application to perform a complete anatomical coronary artery assessment (stenosis grading associated to plaque features analysis) | Good accuracy in plaque phenotype characterization (AUC 0.77) and in determining its anatomical significance (i.e., stenosis degree above or below 50%, AUC 0.80) |
| Han et al. (73) | Automated ML algorithm to identify RPP | The ML model that included clinical variables, qualitative and most importantly quantitative plaque features showed the highest performance in identifying patients at risk of RPP (AUC 0.83) |
| Choi et al. (74) | DL application to perform a complete anatomical coronary artery assessment (stenosis grading associated to plaque features analysis) and CAD-RADS classification | Accuracy compared to three expert readers' analysis for stenoses >70%: 99.7%; accuracy for stenoses>50%: 94.8%. Excellent concordance in CAD-RADS classification with expert readers: agreement within one CAD-RADS category: 98% exams per-patient; 99.9% vessels on a per-vessel basis. |
| AI powered CT-FFR | ||
| Coenen et al. (75) | Definition of the diagnostic accuracy of a ML application to CT-FFR | In the per-vessel analysis, ML-CT-FFR improved diagnostic accuracy by 20% compared to CTA (from 58 to 78%). The per-patient accuracy improved by 14% compared to CTA (from 71 to 85%). Seventy-three percent false-positive CTA results were correctly reclassified by ML-CT-FFR |
| Nous et al. (76) | Feasibility of ML-CT-FFR application in patients with DM | Overall diagnostic accuracy of ML-CT-FFR in diabetic patients was higher (83%) than in non-diabetic patients (75%); AUC 0.88 and 0.82 for diabetic and non-diabetic patients, respectively |
| Baumann et al. (77) | Differences in ML-CT-FFR application between patients of different genders | ML-FFR-CT equally performed in both genders, not showing significative difference in the AUC between males (0.83) and females (0.83) |
| Tesche et al. (78) | Feasibility of ML-CT-FFR application in the presence of heavy calcifications | No statistically significant differences in the diagnostic accuracy, sensitivity, or specificity of ML-CT-FFR were observed across CT scans of patients attributed to different Agatston score categories |
CAD-RADS, Coronary Artery Disease Reporting and Data System; DM, diabetes mellitus; CP, calcified plaque; DS, diameter stenosis; FFR, fractional flow reserve; MLA, minimal luminal area; NCP, non-calcified plaque; NRS, Napkin ring sign; RPP, rapid plaque progression.