Introduction
Assessment of degree of stenosis (DS) in coronary arteries using coronary computed tomography angiography (CCTA) is an accepted diagnostic tool for the detection and exclusion of coronary artery disease (CAD), with consistently high sensitivity and negative predictive value [
1‐
3]. However, it has limited specificity in indicating the functional significance of a stenosis [
2,
3]. Invasive fractional flow reserve (FFR) is currently the reference standard to indicate functional significance of a coronary stenosis and to guide treatment [
4]. However, due to its invasive nature and high cost, adoption of invasive FFR in clinical practice is limited, and the search for a non-invasive method that would determine functional significance of a stenosis continues.
To address the limited specificity of CCTA, new techniques have been developed to obtain information about the functional significance of a stenosis in a non-invasive way. FFR derived from CT (FFRct) is an emerging method which has shown promising results [
5‐
7]. By simulating flow and pressure through the coronary arteries, a virtual FFR value is obtained. More recently, promising results have been obtained with analysis of myocardial perfusion from resting CT [
8‐
16]. Even though it is well known that perfusion defects are more pronounced under conditions of hyperemia [
17,
18], prior studies have shown the feasibility and accuracy of identification of patients with a functionally significant coronary artery stenosis with resting CCTA only [
8‐
13]. With these approaches, functional information is obtained without the need for an additional stress perfusion acquisition, thereby saving radiation and contrast medium dose, lowering risk, and reducing examination duration and cost. In recent studies, approaches exploiting machine learning have been proposed in which the left ventricular myocardium (LVM) in resting CCTA is analyzed and used to classify patients with regard to the presence of functionally significant coronary artery stenosis [
8,
9,
13].
In classical machine learning, discriminant features describing the LVM, such as hypo-attenuation and changes in myocardial wall thickness, are manually designed by an expert. Subsequently, these features are used in an algorithm that is built to classify patients according to presence of functionally significant coronary artery stenosis [
8,
9]. In contrast to the approaches using expert engineered image features, we recently proposed a deep learning (DL) algorithm, whereby the LVM features that discriminate patients with and without functionally significant coronary artery stenosis are independently learned by the algorithm directly from the image [
13]. The current study expands on our previous work [
13] by applying a combined method of visual stenosis grading on CCTA and only applying the DL-based analysis to the intermediate-degree stenosis. In contrast to classical machine learning–based approaches, the DL algorithm is able to independently learn generic and complex LVM patterns, and could potentially be more sensitive to changes in the LVM caused by functionally significant stenosis [
19,
20].
Thus, the aim of the current study was to evaluate the added value of resting CCTA LVM deep learning analysis over coronary DS evaluation only, for identification of patients with functionally significant coronary artery stenosis.
Discussion
In the current study, we aimed to evaluate the added value of resting CCTA LVM deep learning analysis over coronary DS evaluation only, for identification of patients with functionally significant coronary artery stenosis. We demonstrated that the combination of DS with DL analysis of the LVM in intermediate-degree coronary stenosis improves the specificity of CCTA in comparison to evaluation of the DS alone.
Visual determination of the DS from CCTA is a highly sensitive and established diagnostic tool for evaluation of patients with chest pain of suspected coronary origin. However, CCTA presently lacks specificity for identifying the functional significance of coronary stenosis [
1,
2]. A tiered approach with subsequent functional testing (e.g., stress myocardial perfusion imaging) can reduce the number of patients undergoing ICA [
26]. However, this approach mandates additional diagnostic testing before deciding on the necessity for PCI. Here, we propose combining evaluation of DS with LVM DL analysis to improve the specificity of CCTA. The combined evaluation has the potential to avoid additional examinations because analysis is performed on the already acquired CCTA images.
Analysis of the LVM on resting CCTA has been previously studied. Two studies evaluated resting dual-energy CT for the detection of perfusion defects confirmed by ≥ 50% stenosis at ICA [
15,
16]. They found a slight decrease in sensitivity (79–90%) and increase in specificity (86–92%) compared to ≥ 50% DS on CCTA only (82–98% and 88–91%, respectively). Osawa et al [
10] visually evaluated perfusion of the LVM in all cardiac phases from a retrospectively ECG-triggered scan at rest and compared their results with invasive FFR measurements. By combining this LVM analysis with DS, they found an incremental value (AUC = 0.82) over CCTA alone (AUC = 0.71), which is in line with our study. In two recent studies, machine learning–approaches using expert-designed features for classifying patients with functionally significant stenosis were described [
8,
9]. Xiong et al [
8] found a good discrimination for the detection of a perfusion defect (max. AUC = 0.73). However, in this study, a stenosis degree of ≥ 50% DS on ICA was used as reference [
8], and no FFR was performed, making a direct comparison with our results impossible. Han et al [
9] used FFR as reference and described that their algorithm, combined with evaluation of DS, showed an added value (AUC = 0.75) over DS alone (AUC = 0.68) [
9], which is closely in line with the findings in the present study. However, patients with intermediate DS comprised a minority of the subjects studied by Han et al [
9] (33%, 82/252), whereas the present study consisted of a majority of intermediate stenosis (80%, 101/126) and was exclusively focused on the added value of DL in this most challenging patient category.
FFRct is another technique performed on resting CCTA that has shown promising results (AUC = 0.79–0.93) for the evaluation of functionally significant stenosis [
5‐
7]. However, FFRct depends on lumen segmentations which can be challenging or impossible in patients with high-density calcified plaque, motion or misalignment artifacts, and/or prior CABG or PCI [
5‐
7]. Our proposed method may be less affected by these challenges as the analysis of intermediate stenosis is performed only on the LV myocardium. Results of the present study underscore the need to look beyond the coronary arteries in the quest to improve specificity of CCTA. A recent systematic review by Cook et al [
27] found that FFRct has a lower diagnostic performance with FFRct values around the cut point (0.7–0.8). In addition, patients with physiologically intermediate lesions (invasive FFR 0.7–0.8) comprised only a small minority of all the patients studied (12.8%) with a median FFR of 0.88, indicating a focus on patients with milder disease. In the current study, nearly three times as many patients (36.5%, 46/126) had physiologically intermediate lesions, and median FFR was 0.78, indicating a more diseased population. This supports the notion that the value of additional analysis of the myocardium is likely to be highest in patients with intermediate stenosis. However, a limitation of the present study is that results are only available on a patient basis; therefore, no indication could be given as to which stenosis was functionally significant, while FFRct has the ability to evaluate functional significance of a stenosis on a vessel basis. It is likely that a combined approach using both techniques will lead to further increase in specificity for identifying flow-limiting lesions in patients with intermediate-degree stenosis at CCTA.
DL algorithms are able to learn new complex patterns and relationships directly from the images without prior specification of which image features represent presence of disease. Therefore, DL may be more sensitive to subtle changes in the LVM caused by functionally significant stenosis, which can be difficult to detect by a human observer [
19,
20]. In spite of excellent performance of deep learning techniques demonstrated in many medical image analysis tasks, their interpretability is very limited, and we do not yet fully understand their inner working [
19,
28]. Also in this work, even though the extracted encodings are relevant in representing LVM (Supplement Fig.
1), they are not readily interpretable. As distinct combinations of encodings represent different LVM appearances (e.g., normal vs. thin), specific encodings do not correspond to specific physical appearances. This prevented us from interpreting, visualizing, and localizing differences within the LVM for patients with significant stenosis. Future work might address these limitations [
19,
28].
This study has limitations. First, the retrospective single-center study design has to be taken into account. This could have induced selection bias as CT could have been the reason for referral to invasive FFR. In addition, because of this bias, the study group consisted of patients with high Agatston scores and a high prevalence of stenosis and extensive CAD with mainly intermediate stenosis (
n = 101/126). Due to blooming, calcified plaques are known to cause overestimation of DS and thereby decrease specificity [
29]. In addition, as reported in prior studies [
2,
3], DS on CCTA showed high diagnostic accuracy for low- and high-grade stenosis (≤ 24% and ≥ 70%), which decreased when evaluating intermediate-degree stenosis (25–69%).This is reflected in the low specificity for DS only on CCTA found in the current study (31.1%) compared to reported in literature (40–83%) [
1‐
3]. Second, maximum DS for each coronary artery was visually categorized, and no continuous quantitative measurements were performed. Although continuous quantitative measurements would allow for a more accurate way to evaluate diagnostic accuracy of DS, this is not regularly performed in clinical practice and was therefore not performed. Third, in the current study, no comparison with another type of functional assessment was performed (e.g., FFRct or visual assessment of myocardial perfusion), thereby limiting the evaluation of the added value of the current method compared to other available methods. Fourth, we used equipment of one single vendor and results may therefore be limited to this vendor. A final limitation is the relative small patient cohort and unbalanced dataset, which can introduce bias in performance. Although we performed 50 repetitions of 10-fold cross-validation experiments with randomization after each repetition, results may still be partially affected by coincidental findings. Analysis on a separate patient cohort as well as prospective studies needs to be performed to assess whether the found correlation also implies causation. Future work will address these limitations.
In conclusion, combining assessment of degree of stenosis with DL analysis of the LVM may result in improved diagnostic performance for identification of patients with functionally significant coronary artery stenosis. Future research is warranted to evaluate this approach on patients with low pre-test probability of obstructive coronary disease, more clinically encountered at CCTA.
Compliance with ethical standards