Cardiac CT: Technological Advances in Hardware, Software, and Machine Learning Applications

Abstract

Purpose of Review

Multidetector row computed tomography (CT) allows noninvasive imaging of the heart and coronary arteries. The purpose of this review is to briefly summarize recent advances in CT hardware and software technology, and machine learning applications for cardiovascular imaging.

Recent Findings

In the last decades, there have been significant improvements in CT hardware focusing on faster gantry rotation resulting in improved temporal resolution. Concurrent hardware improvements include improved spatial resolution and higher coverage of the patient, enabling faster acquisition. Advances in cardiac CT software include methods for measurement of noninvasive FFR, coronary plaque characterization, and adipose tissue characteristics around the heart. Machine learning approaches using cardiac CT have been shown to improve both risk of prognosis and lesion-specific ischemia.

Summary

Recent advances in CT hardware and software have expanded the clinical utility of CT for cardiovascular imaging. In the next decades, continued advances can be anticipated in these areas, and in machine learning applications in cardiac CT, as they are incorporated into clinical routine for image acquisition, image analysis, and prediction of patient outcomes.

Introduction

In recent years, there have been significant advances in CT hardware, software, and machine learning, which have expanded the clinical utility of CT for cardiovascular imaging. Imaging of the coronary arteries is challenging due to their small dimensions, tortuosity, and continuous motion. Cardiovascular CT needs full synchronization with the electrocardiogram (ECG) signal and high temporal resolution to “freeze” cardiac motion, and has benefitted from steadily increasing gantry rotation speeds. Imaging of the coronary arteries also requires high, near-isotropic spatial resolution. Multiple cross-sections (ranging from 64 to 320 with corresponding through-plane or z-coverage of 4–16 cm) need to be acquired simultaneously for whole heart acquisition in a single breath-hold. In response to these technical requirements, there have been rapid advances in CT hardware; such advances remain ongoing. Concurrently, CT hardware advances have been accompanied by advances in software to facilitate and extend cardiac CT interpretation. This review summarizes in brief recent technological advances in CT hardware and software, and recent promising machine learning approaches to quantify cardiovascular risk and lesion-specific ischemia.

Current State-of-the-Art in CT Hardware

The key specifications for current state-of-the-art CT scanners from the major manufacturers are listed in Table 1. A general technical improvement is that gantry rotation times have been decreased by the manufacturers to address the requirement of high temporal resolution for clinical cardiac CT. Concurrently, there has been an effort to increase the spatial resolution, as well as the z-coverage to facilitate faster scanning of the heart. For low (< 60 beats/min) and regular heart rates, single heartbeat scans can be performed. Major advances in current CT hardware are highlighted below.

Table 1.

Key specifications for current state-of-the-art CT scanners from the major manufacturers

Scanner	Fastest rotation time (ms)	Temporal resolution (ms)	Detector row width (mm)	Number of detector rows	Detector Z-axis coverage (mm/rotation)
GE Revolution	280	140	0.625	256	160
GE Cardiographe	240	120	0.5	192 (280 slices per rotation)	140
Hitachi SCENARIA	350	175	0.6	64	38.4
Philips Brilliance iCT	270	135	0.625	128	80
Samsung NExCT 7	250	125	0.625	128	80
Siemens FORCE	250	66	0.6	96 (192 slices per rotation)	57.6
Toshiba AquilionONE	275	137.5	0.5	320	160

As listed in Table 1, for single-source CT (single X-ray tube), the temporal resolution is half the gantry rotation time. One manufacturer has developed a dual-source CT scanner with two X-ray tube/detector systems on the same gantry; the temporal resolution for this system is reduced to one fourth of the gantry rotation time (Table 1).

Detectors

Standard multidetector CT (MDCT) detector elements are solid-state scintillators coupled to a photodiode; the role of the scintillator is to convert X-rays to visible light. The photodiode converts visible light into electrical signal which is transmitted to the computer. Recently, new scintillator materials with shorter decay times have been introduced, with improvement in the response time of the detector. Faster detector response times can effectively support increased data sampling per rotation, allowing improved in-plane (x-y) spatial resolution. Using a garnet-based scintillator material, better image quality and accuracy of in vivo coronary stent imaging (with diameters ranging from 2.75 to 3.5 mm) was demonstrated compared to older technology [1, 2].

Fully integrated circuit detectors have also been introduced that combine photodiodes and electronics, reducing electronic noise. Higher spatial resolution with integrated circuit detectors was shown to improve image quality for coronary CT angiography and quantitative assessment of in-stent restenosis in phantom studies [3, 4]. The highest reported in-plane spatial voxel dimensions of current MDCT scanners are in the range of 0.23–0.4 mm and 0.3 mm in the z direction.

Tube Voltage

Till recently, the lowest allowed peak tube voltage in CT scanners was 80 kVp. A 70-kVp setting has been introduced on some CT scanners. Lower tube potential can potentially improve image contrast—due to increased probability of interaction via photoelectric effect at lower energies between incident X-rays and body tissue. Concurrently, the limitation of lower tube voltage is image noise due to lower X-ray penetrability. Lower kVp is indicated for cardiac imaging of pediatric patients and can also be applied to adults with smaller body size, to reduce radiation dose. Hell and colleagues have demonstrated the feasibility of very low radiation dose coronary CT angiography, with 70 kVp and high-pitch prospective ECG-triggering, in 27 patients with body weight < 100 kg and regular heart rate < 60 beats/min [5•]. In another study of 45 patients with high-pitch prospective ECG-triggering, X-ray tube voltage was based on body mass index (BMI): 80 or 70 kV for BMI < 26 kg/m² versus 100 kV for 26–30 kg/m² [6]. This study showed that in nonobese patients, use of 70 kVp and high-pitch prospective ECG-triggering mode results in robust image quality for the coronary arteries at reduced radiation dose and contrast volume.

Dual-energy CT acquisition refers to CT acquisition at two different energies (high and low), which yields two distinct CT attenuation data sets; the differences of X-ray attenuation between high and low energies can improve tissue characterization. The highest changes between low and high energies are in more dense tissues such as iodine and calcium; signals from less dense tissues such as noncalcified plaque and fat show less change [7]. Currently, dual-energy acquisition uses either two X-ray tube/detector systems (for example, with one tube at 140 kVp and the other at 120 kVp), single X-ray tube/detector system with rapid tube voltage switching, a single X-ray tube with a filter for X-ray beam energy splitting, single X-ray tube and dual layers of energy-sensitive detectors, or a single X-ray tube with a wide detector array. Processing of dual-energy CT data yields several sets of three-dimensional datasets. For example, since iodine is identified from the dual-energy data, blood pool images with only iodinated contrast can be obtained. Additionally, images with no iodine (called virtual noncontrast images) can be obtained. To date, the majority of CTA scans are not performed in dual-energy mode; however, dual-energy CT has been applied for several cardiovascular indications and particularly for CT myocardial per-fusion imaging [8–11]. In initial studies with vascular imaging, dual-energy CT has been shown to improve separation of calcified plaque from noncalcified plaque and lumen components [12–15], and to improve assessment of stents and stent patency [16]. In a study with virtual histology intravascular ultrasound, dual-energy CT has been shown to be increase the accuracy of detection of necrotic core in 7 postmortem coronary arteries and 20 patients [17].

Iterative Reconstruction

Iterative reconstruction has been extensively used over the last decades in nuclear cardiac perfusion imaging, where the reconstructed image sizes are much smaller. In recent years, iterative reconstruction has been gradually replacing standard filtered back projection in cardiac CT. Most currently available commercial implementations model photon counting statistics or noise and allow reduction of noise and improved delineation of edges [5•, 18–21]. Some commercially available algorithms further include models of the scanner geometry and X-ray interaction with matter in addition to noise and are grouped under the general term “model-based” [22–24]. These “model-based” methods are the most computationally intensive but allow the highest accuracy and improvement in spatial resolution, as well as noise reduction. With the increase of the computing capacity of standard computers, such advanced model-based reconstruction techniques which improve image quality, accuracy, and spatial resolution are likely to be implemented.

Iterative reconstruction methods have been primarily used to significantly decrease the radiation dose to the patient without compromising image quality. In these “reduced-photon” data, typically, a combination of these techniques was used: decrease in X-ray tube current/voltage and increased pitch (high-pitch prospectively ECG-triggered acquisition mode) [5•, 19, 22, 24]. The large, prospective, multicenter, multivendor PROTECTION V trial has assessed the potential of iterative reconstruction for radiation dose reduction in coronary CTA [25]. Four hundred consecutive patients were randomized to one of two groups: (i) standard filtered back projection with standard tube current, or (ii) to iterative reconstruction with 30% tube current reduction. This trial showed that coronary CTA image quality is maintained with the combined use of 30% reduced tube current and iterative reconstruction when compared with conventional filtered back projection and standard tube current [25].

Coronary Calcium Scoring

Coronary calcium scoring with noncontrast CT is used worldwide for cardiovascular risk stratification and provides a measure of coronary atherosclerotic burden. It is a simple and low-cost test, measuring coronary artery calcium without use of pre-medication or contrast, and with a uniformly low attendant radiation burden, acquired with prospective ECG-triggering [26–28]. To date, several follow-up studies have reported that the total coronary artery calcium measured by noncontrast CT predicts cardiovascular events [29–43], beyond standard cardiovascular risk indices.

Coronary CT Angiography

Coronary CT angiography (CTA) has recently emerged as a noninvasive diagnostic test in selected stable but symptomatic patients with suspected CAD [44–49]. Current clinical interpretation of CTA relies on visual assessment of stenosis grade, plaque type, and presence of high-risk plaque features [50••, 51••]. Invasive coronary angiography is the gold standard for the diagnosis of obstructive stenosis. CTA has, to date, shown high accuracy and sensitivity for detecting coronary artery stenoses when compared to invasive coronary angiography [45–49]. In particular, CTA has shown very high negative predictive value for obstructive stenoses (range 89–99% in multicenter studies) in patients with symptoms that suggest the presence of coronary artery disease [45–49].

Figure 1 shows an example of such a negative CTA with normal coronary arteries acquired at low radiation dose with prospective ECG-triggering in a patient with suspected coronary artery disease. Currently, CTA data can be acquired with prospective ECG-triggering of X-rays (with axial or helical high-pitch scan mode) or retrospective ECG-gating (with helical scan mode), depending on patient heart rate, arrhythmia, and other factors [52]. In the prospective, multicenter and multivendor PROTECTION III trial (randomized study of 400 patients), prospective ECG-triggered axial scanning reduced radiation exposure by 69% compared to retrospective ECG-gated helical scanning, with equivalent image quality [53]. Thus, in patients with low and stable heart rates, prospective ECG-triggered scan is recommended.

Fig. 1.

Example of a normal CTA scan (no stenosis or plaque in any coronary artery) in a 43-year-old woman with family history of coronary artery disease. Heart rate was 65 beats/min for CTA. Effective dose for this axial prospectively ECG-triggered scan acquired with 100 kVp, was 1.5 mSv [dose length product (DLP) 107.1 mGy cm]

The international CONFIRM registry (with > 25,000 patients), as well as the large, prospective randomized PROMISE trial (with 10,000 patients), has indicated that if CTA performed to detect coronary stenosis is “negative”, no further testing is necessary [54, 55••]. In the prospective, randomized PROMISE trial with 10,003 patients, performance of CTA was equivalent to stress testing when used as an initial test for suspected coronary artery disease, with similar event rates between stress testing and the CTA group at 2-year follow-up (3.3 vs. 3.0%) [55••]. While the utilization rate of invasive coronary angiography was higher in patients with CTA as the initial test (12.2 vs. 8.3%), prevalence of nonobstructive stenoses was significantly higher in the population with stress testing (52.5%) than in the CTA-tested group (27.9%), demonstrating the high negative predictive value of CTA.

In the prospective, randomized SCOT-HEART trial, 4146 patients with suspected coronary artery disease were randomly assigned to standard care alone (typically, stress testing) or standard care plus CTA [56••]. After 1.7 years’ follow-up, CTA was associated with a 38% reduction in myocardial infarction (fatal and nonfatal, with a trend towards significance p = 0.0527). In this trial, CTA was shown to clarify the diagnosis, enable appropriate treatment, and improve patient outcomes.

Cardiac CT Software

Beyond stenosis, several advances have been made in software for standard cardiac CT scans, in the areas of quantitative plaque characterization, measurement of noninvasive fractional flow reserve (FFR), and epicardial adipose tissue. These are briefly summarized below.

Plaque

Beyond stenosis, CTA also permits assessment of atherosclerotic plaque (including plaque burden and composition) and outward coronary artery remodeling, previously only measurable through invasive means [57–61]. Studies have shown that coronary plaque volume and remodeling quantified from CTA correlate strongly with invasive intravascular ultrasound (IVUS) [59, 62, 63, 64•, 65, 66]. Features of plaque vulnerability measured by CT have been reported to be low-attenuation or low-density plaque—with attenuation values ≤ 30 Hounsfield units (HU)—corresponding to the necrotic core, and positive or outward remodeling [61, 67, 68, 69••, 70, 71••]. The napkin-ring sign has also been shown to be a CT signature of high-risk coronary atherosclerotic plaque [69••, 72••, 73]. A key long-term study by Motoyama et al. has investigated whether visually assessed plaque characteristics by CTA can predict mid-term likelihood of future ACS over mean 3.9 years of follow-up [71••]. The presence of LAP, positive remodeling, and severe stenosis by CTA (≥ 70%)—defined as high-risk plaque features—were visually assessed in 3158 patients undergoing CTA. ACS occurred in 88 patients and was significantly more frequent in patients with severe stenosis. CTA-verified high-risk plaque was an independent predictor of ACS. However, the cumulative number of patients suffering ACS was similar for patients with and without visually identified high-risk plaques [71••]; this was primarily attributed to the diffuse nature of atherosclerosis. Notably, quantitative plaque analysis was not performed in this study. Plaque progression, visually defined by an increase in stenosis grade from serial CTA, was found to be an independent predictor of ACS [71••].

In studies with comprehensive quantitative plaque analysis with semi-automated software, low-density noncalcified plaque (NCP) burden has been shown to improve prediction of future adverse cardiac events, even in diffuse atherosclerosis [74, 75••]. Figure 2 shows an example of semi-automated quantification of high-risk coronary plaque, with large NCP and low-density NCP burden and coronary artery remodeling. Hell et al. have investigated whether quantitative global per-patient plaque characteristics measured by coronary CTA can predict subsequent cardiac death, during 5-year midterm follow-up. In this study, low-density noncalcified plaque, noncalcified plaque (NCP), and total plaque volumes, as well as contrast density difference (a measure of luminal contrast kinetics), were the strongest independent predictors of future cardiac death, even when adjusted for Segment Involvement Score [75••]. Semi-automated plaque quantification can also allow measurement of individualized plaque progression or regression. The effect of lipid-lowering therapy with statins on quantitative change in coronary plaque burden [76, 77] and on plaque composition [78, 79] has been explored. Intensive low-density lipoprotein (LDL) cholesterol lowering has been shown to attenuate total plaque progression in a multicenter study of 467 patients who underwent serial CTA 2 years or more apart [80]. In a very recent study with quantitative plaque characterization in serial CTA scans (116 patients), LDL cholesterol reduction by as low as ≥ 10% (of baseline LDL-values) was characterized by beneficial changes in amount and composition of plaque, with a significant decrease in NCP and total plaque volumes, and no significant change in calcified plaque volumes [81], thus providing mechanistic insight into the prognostic benefits of lipid-lowering therapy.

Fig. 2.

High-risk coronary plaque quantification in the proximal left anterior descending artery by semi-automated software. a Curved MPR and straightened views showing NCP in red and CP in yellow. b 3D arterial view with arrow showing low-density noncalcified plaque (surrogate marker for the lipid core)

Revascularization strategies guided by lesion-specific ischemia (characterized by invasive FFR values ≤ 0.80) have been shown to improve patient outcomes [82, 83]. Low-density NCP —which has been shown to be of prognostic importance—has also been shown to be an imaging biomarker which predicts lesion-specific ischemia by invasive FFR [84••, 85]. In the multicenter NXT trial, low-density NCP provided independent and incremental discrimination of ischemia beyond stenosis severity [84••].

Hemodynamically Significant Stenosis

It is known that stenosis does not equate ischemia, and significant stenosis in CTA often does not indicate a hemodynamically significant lesion. Although coronary CTA is an anatomic test, in recent years, computational fluid dynamics (CFD)-based methods from anatomical CTA have been developed to simulate FFR. Noninvasive FFR (FFR_CT), derived from standard CTA based on CFD-based methods from CTA, have been reported to be superior to CTA anatomical interpretation in prospective multicenter studies [86–88]. In the largest of these studies (the “NXT” trial), overall per-patient sensitivity, specificity, and accuracy of FFR_CT to identify hemodynamically significant lesions (with FFR ≤ 0.80) were 86, 79, and 81% [86]. Figure 3 shows an example of FFR_CTand diffuse plaque measured in the LAD artery of a 52-year-old male CTA patient with cardiovascular risk factors.

Fig. 3.

a FFR-CT in left anterior descending artery with nonobstructive stenosis for a 52-year-old male with cardiovascular risk factors, showing progressively decreasing noninvasive FFR. b Quantitative plaque analysis showed diffuse atherosclerotic plaque with extensive noncalcified plaque (760 mm³, red), calcified plaque (67 mm³, yellow), maximum contrast density difference of 30% and maximum diameter stenosis of 47%

Onsite FFR_CT using a reduced-order CFD computation on a research workstation has also been reported [89]; recently, this simulation has been accelerated by machine learning [90].

CT perfusion has also been shown to effectively predict the hemodynamic significance of coronary stenoses [91, 92]. In the multicenter Core 320 trial, integrated coronary CTA and CT perfusion could effectively predict flow-limiting coronary artery stenosis assessed by invasive coronary angiography and single photon emission computer tomography (AUC 0.87) [92]. CT perfusion requires an additional stress CT scan, with attendant radiation exposure and need for pharmacologic stress.

Epicardial and Pericoronary Adipose Tissue

Epicardial adipose tissue (EAT) is a metabolically active fat depot encompassed by the pericardium, and directly surrounding the coronary arteries. Epicardial adipose tissue characteristics (volume and density) have been measured from cardiac CT, from both noncontrast coronary calcium scoring scans, as well as CTA, using semi-automated software. There is increasing evidence that EAT exerts a local pathogenic effect on coronary vasculature and on the heart [93–97] and is related to adverse cardiovascular events [97, 98].

In a recent study of 456 asymptomatic patients by Goeller et al., lower EAT density and increased EAT volume measured from coronary calcium scoring CT using semi-automated software (Fig. 4) were associated with serum levels of plaque inflammatory markers and major adverse cardiovascular events, suggesting that dysfunctional EAT may be linked to early plaque formation and inflammation [99••]. EAT density was more significantly related to events (hazards ratio 0.8, p = 0.029) than EAT volume or the coronary calcium score.

Fig. 4.

Epicardial adipose tissue quantification of a 63-year-old asymptomatic female (BMI 25 kg/m²) with family history of CAD, an EAT volume of 50 cm³, mean EAT density of − 79 HU, and a coronary calcium score (CCS) of 154. EAT is highlighted in purple color

Pericoronary adipose tissue, immediately surrounding the coronary arteries, is known to be related to coronary atherosclerosis [100]. A recent study has shown that pericoronary adipose tissue density and its gradient measured from CTA are related to tissue inflammation by histology from 453 patients undergoing cardiac surgery [101••].

Machine Learning Applications in Cardiac CT

Machine learning denotes a category of computer algorithms which learn rules and identify patterns progressively from large datasets, without making any prior assumptions. Machine learning techniques have been effectively used for prediction and intelligent decision-making in many areas of everyday living [102, 103]. In cardiac imaging, there has been two major approaches. In the first (reported more frequently), machine learning has been utilized to predict diagnostic or prognostic outcomes from large datasets with a multitude of clinical and imaging variables. In the second category are a new class of machine learning algorithms, deep learning, with convolutional neural networks, which allow for direct analysis of extracted images for segmentation or outcome prediction.

The feasibility and accuracy of machine learning to predict all-cause mortality at 5-year follow-up was evaluated in the CONFIRM registry (10,030 patients). All available clinical and visually assessed CTA measures were objectively evaluated. Machine learning risk score combining clinical and CTA data exhibited a significantly higher AUC (0.79) for the prediction of death, compared to established risk indices and visual CTA assessment [104••]. For the Multi-Ethnic Study of Atherosclerosis (MESA) study (> 6800 asymptomatic patients undergoing CAC scoring), machine learning demonstrated superior performance than the Coronary Calcium score to predict adverse cardiovascular events [105].

Machine learning has also been applied to predict lesion-specific ischemia. In the NXT trial (254 patients), machine learning combination of clinical data and quantitative stenosis and plaque features exhibited higher AUC compared to pre-test likelihood of coronary artery disease or quantitative CTA metrics for predicting ischemia (ML 0.84 vs. best clinical score 0.63, CTA stenosis 0.76, low-density noncalcified plaque volume 0.77, p < 0.006) [106••] (Fig. 5). Contrast density difference had the highest information gain to identify lesion-specific ischemia [106••].

Fig. 5.

Prediction of lesion-specific ischemia by the integrated ischemia risk score by machine learning (ML-combined) and plaque metrics. a ML-combined versus quantitative plaque volumes (LD-NCP, NCP and total plaque volume). b Ml-combined versus quantitative stenosis and pre-test likelihood of coronary artery disease. ML-combined had a significantly higher AUC compared to individual quantitative CTA plaque measures or the pre-test likelihood. Reproduced from Dey et al [106••] (with permission)

Finally, machine learning has also been applied to estimate CTA image quality with similar results as expert visual assessment [107].

Deep learning, using convolutional neural networks, has been utilized for automated coronary calcium scoring from low-dose CT and CTA, showing good agreement with the expert reader [108, 109]. Automated quantification of EAT from calcium scoring CT scans using convolutional neural networks has been investigated on a study with 250 patients; in this study, fully automated quantification showed high agreement and correlation to expert manual measurement [110] (R = 0.924, p < 0.00001). This technique can also be used for image-based identification of disease. In a recent study, deep learning was used for automated analysis of the myocardium in CTA to identify coronary stenoses with hemo-dynamic significance [111]. Deep learning has also enabled faster direct computation of noninvasive FFR. The method was trained on a large database of synthetically generated coronary models and shown to be equivalent to an onsite computational fluid dynamics-based algorithm, with shorter execution times [90, 112].

Conclusion

In conclusion, over the last few years, cardiac CT imaging has demonstrated continuous progress in CT hardware, software, and machine learning applications. For CT hardware, significant advances have been made in scanner gantry rotation speeds improving temporal resolution, spatial resolution, detector coverage, and therefore overall scan times and needed contrast. In the next decades, we can anticipate continued progress in all these areas, and particularly in machine learning applications in cardiac CT, as they are incorporated into clinical routine for image acquisition, image analysis, and prediction of patient outcomes.