Cardiac CT in the era of artificial intelligence: precision imaging, treatment guidance and optimised risk stratification for coronary artery disease

Figure 1. Overview of AI-enhanced cardiac CT in coronary artery disease. AI, artificial intelligence; CACS, coronary artery calcium scores; CAD-RADS, coronary artery disease reporting and data system; CT-FFR, CT-derived fractional flow reserve; CTO, chronic total occlusion; EAT, epicardial adipose tissue; DL, deep learning; ECV, extracellular volume; ICA, invasive coronary angiography; LIE, late iodine enhancement; MBF, myocardial blood flow; ML, machine learning; PCAT, pericoronary adipose tissue; PCI, percutaneous coronary intervention; SYNTAX, synergy between percutaneous coronary intervention with taxus and cardiac surgery.

Figure 2. Common issues during coronary artery reconstruction. (A, B) Green arrows indicate motion artefacts from respiration or cardiac pulsation, causing blurred segments in the curved planar reformation (CPR) images and discontinuities in the segmented coronary artery tree. (C, D) Green arrows indicate veins with morphology and location similar to those of coronary arteries that may be misidentified as arteries, creating false segments in segmentation results. Schematic of manual and automated CAC measurement, where coloured regions represent calcifications from different coronary artery branches: (E) manual CAC annotation based on ECG-gated non-contrast cardiac CT; (F) DL-based CAC annotation from nongated images (eg, a single CT-MPI phase) yields results comparable to manual annotations from gated scans; (G, H) DL tools automatically identify CAC even in lower-quality images. CAC, coronary artery calcium; CT-MPI, CT myocardial perfusion imaging; DL, deep learning.

Figure 3. Schematic of DL-based CT-FFR and examples of LIE and ECV imaging. (A) Entire coronary artery tree. (B–D) Branch-specific quantification, with CT-FFR values labelled numerically at the indicated sites. Examples of LIE and ECV images: (E, F) CCTA LIE images before and after denoising; (G, H) ECV maps derived from LIE images showing a normal patient (mean<30) and an abnormal patient (mean>30). CCTA, coronary CT angiography; CT-FFR, CT-derived fractional flow reserve; DL, deep learning; ECV: extracellular volume; LIE, late iodine enhancement.

Figure 4. An overview of how CT-MPI data is processed by DL. (A–E) CT-MPI acquires~10–15 phases during image acquisition; cardiac motion causes position variation across phases. Image registration is required before analysis. DL tools can directly map raw data to final parametric maps. (F, G) MBF short-axis pseudocolor map and 17-segment bull’s-eye plot of a normal patient; (H, I) MBF short-axis pseudocolor map and 17-segment bull’s-eye plot of a patient with ischaemia (dark green area). CT-MPI, CT myocardial perfusion imaging; DL, deep learning; MBF, myocardial blood flow.

Figure 5. Comparison of manual and DL–based coronary CTO reconstruction. (A) Case 1: (from left to right) manual reconstruction before editing shows the failure of the commercial workstation to extract the occluded segment (dashed lines); manual reconstruction after editing with the occluded segment manually added using a multipoint vessel-tracking tool (arrow); DL-based reconstruction (MIP) showing automatic extraction of the entire CTO pathway (arrow) with more small branches; and DL-based CPR image showing the reconstructed CTO. (B) Case 2: same sequence of panels as in (A), illustrating another representative case. CPR, curved-planar reformation; CTO, chronic total occlusion; DL, deep learning; MIP, maximum intensity projection.