Evaluation of the diagnostic value of joint PET myocardial perfusion and metabolic imaging for vascular stenosis in patients with obstructive coronary artery disease

To investigate the diagnostic value of joint PET myocardial perfusion and metabolic imaging for vascular stenosis in patients with suspected obstructive coronary artery disease (CAD). Eighty-eight patients (53 and 35 applied for training and validation, respectively) with suspected obstructive CAD were referred to 13N-NH3 PET/CT myocardial perfusion imaging (MPI) and 18F-FDG PET/CT myocardial metabolic imaging (MMI) with available coronary angiography for analysis. One semi-quantitative indicator summed rest score (SRS) and five quantitative indicators, namely, perfusion defect extent (EXT), total perfusion deficit (TPD), myocardial blood flow (MBF), scar degree (SCR), and metabolism-perfusion mismatch (MIS), were extracted from the PET rest MPI and MMI scans. Different combinations of indicators and seven machine learning methods were used to construct diagnostic models. Diagnostic performance was evaluated using the sum of four metrics (noted as sumScore), namely, area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. In univariate analysis, MIS outperformed other individual indicators in terms of sumScore (2.816-3.042 vs 2.138-2.908). In multivariate analysis, support vector machine (SVM) consisting of three indicators (MBF, SCR, and MIS) achieved the best performance (AUC 0.856, accuracy 0.810, sensitivity 0.838, specificity 0.757, and sumScore 3.261). This model consistently achieved significantly higher AUC compared with the SRS method for four specific subgroups (0.897, 0.839, 0.875, and 0.949 vs 0.775, 0.606, 0.713, and 0.744; P = 0.041, 0.005, 0.034 0.003, respectively). The joint evaluation of PET rest MPI and MMI could improve the diagnostic performance for obstructive CAD. The multivariate model (MBF, SCR, and MIS) combined with SVM outperformed other methods.


INTRODUCTION
Coronary artery disease (CAD) refers to coronary artery atherosclerotic lesions that cause stenosis or vascular lumen obstruction, resulting in heart disease caused by myocardial ischemia, hypoxia, or necrosis. 1 CAD was the second leading cause of chronic diseases in China until 2017. 2 Thus, effective and accurate diagnosis is particularly important for the management of patients with suspected or known CAD. Obstructive CAD diagnosis is mainly based on coronary stenosis detection. Coronary angiography (CAG) can effectively determine the presence and degree of coronary stenosis and is the gold standard for CAD diagnosis. 3 However, CAG is a costly, invasive procedure with risk for complications that may be life-threatening and cause irreversible damage. 4,5 Furthermore, CAG is contraindicated in patients allergic to contrast media and with liver and kidney dysfunction because of contrast medium use. 6 Positron emission tomography myocardial perfusion imaging (PET MPI) has higher accuracy, sensitivity, and specificity for CAD detection than single-photon emission computed tomography (SPECT) MPI. [7][8][9][10] Coronary artery calcium (CAC) is a global marker of atherosclerosis, 11 and regional CAC scores improve the accuracy in CAD detection. 12 Ischemic total perfusion deficit (ITPD), which is a measure of hypoperfusion change between stress and rest in the entire ventricle, 13 has been combined with regional and global CAC to enhance the overall diagnostic value of PET/CT for obstructive CAD detection. 14 The combination of CAC score, ITPD, and quantitative coronary vascular function further improves the diagnostic accuracy of 82 Rb PET/CT in predicting obstructive CAD. 15 However, these studies require rest and stress PET MPI.
For CAD diagnosis, pharmacological and exercise stress PET MPI are more sensitive and have higher clinical value than resting PET MPI and have been used in the clinic. However, stress PET is not widely applied in the clinics in China because patients that undergo cardiac PET imaging, mostly those with moderate to severe CAD and especially the elderly, are prone to cardiovascular dysfunctions, such as fatal arrhythmias and serious cardiac malignant events, during stress tests. Hence, doctors and patients are under tremendous psychological stress and risk. The clinical routine management involves performing myocardial metabolic imaging (MMI) PET, followed by a rest PET MPI. MMI reveals scar degree (SCR) and is the gold standard for evaluating the presence and extent of viable myocardium. 16,17 MPI can provide summed rest score (SRS), perfusion defect extent (EXT), total perfusion deficit (TPD), and myocardial blood flow (MBF). 18 A joint evaluation of MMI and MPI PET can provide ''metabolism-perfusion mismatch'' (MIS), which reflects myocardial tissue showing local defects in MPI with no abnormality in the corresponding region of MMI. 19 Therefore, we evaluated the joint value of PET rest MPI and MMI quantification in predicting obstructive CAD. The semi-quantitative SRS and the five quantitative indicators (EXT, TPD, MBF, SCR, and MIS) were combined with seven machine learning (ML) algorithms to derive the optimal combination model and classification method.

Study Population
This retrospective study was approved by the Institutional Review Board, and informed consent was waived. This study initially included 159 patients with suspected or known CAD that underwent 13 N-ammonia ( 13 N-NH 3 ) PET/CT MPI and 18 F-fluorodeoxyglucose ( 18 F-FDG) PET/ CT MMI in rest between October 2017 and June 2019 in Guangdong Provincial People's Hospital, Guangzhou, Guangdong, China. However, 22 patients were excluded due to the following reasons: (1) having severe valvular heart disease, acute myocarditis, uncontrolled arrhythmias, or (2) poor image quality and incomplete clinical information. Among the 137 remaining patients, 88 referred by a clinical physician to CAG within 1 month of PET/CT imaging were finally considered for subsequent analysis.

PET/CT Imaging
All patients separately underwent rest 13 N-NH 3 PET/CT MPI and 18 F-FDG PET/CT MMI scanning on a whole-body Siemens Biography 16 PET/CT scanner in the next 2 days. As a routine preparation for MMI and MPI, the patients were requested to fast for 6 hours and discontinue taking caffeine-containing drugs for 12 hours before examination. The patients were given 50 g of oral glucose and 3 IU insulin load when their fasting blood glucose was B 8.4 mmol/L for MMI scanning. Rest 13 N-NH 3 PET/CT MPI and 18 F-FDG PET/CT MMI scanning protocols were conducted as follows. Following a CT scout acquisition (120 kVp, 10 mA) for patient positioning, a CT transmission scan was obtained (140 kVp, 80 mA) for subsequent attenuation correction. The patients were instructed to breathe normally during PET acquisition. Afterward, 555-925 MBq (15-25 mCi) of 13 N-NH 3 and 18 F-FDG were injected intravenously for MPI and MMI, respectively. A 20-minute dynamic acquisition PET study was obtained. Rest MMI and MPI dynamic images were reconstructed into 21 time frames (12 9 10, 6 9 30, 2 9 60, and 1 9 180 seconds, 10 minutes) after a delay of 180 seconds by using attenuation-weighted ordered subset expectation maximization (two iterations, 24 subsets) and a Gaussian filter (FWHM = 5 mm). CT-based attenuation, scatter, decay, and random corrections were applied to the reconstructed images.

PET Quantitative Image Analysis
Transaxial PET images were automatically reoriented into short-axis and vertical and horizontal longaxis slices. Polar maps of myocardial perfusion and metabolism were generated according to the 17-segment American Heart Association model. 18 The commercially available QPS/QGS software, version 3.0 (Cedars-Sinai Medical Center, Los Angeles, CA, USA) was used to calculate the regional quantitative indicators for each vascular territory: left anterior descending coronary artery (LAD), left circumflex coronary artery (LCx), and right coronary artery (RCA). Quantitative indicators were divided into three categories based on data sources. (i) Perfusion: EXT, TPD, and MBF were calculated from perfusion data; EXT and TPD are percentages indicating the extent and degree of perfusion defect, respectively. MBF is a continuous value of each vascular territory, representing the volume of blood flow through a unit mass of myocardium in a unit time (mL/min/g), and it was computed from the dynamic rest myocardial perfusion imaging series. (ii) Metabolic: after myocardial infarction, the infarcts were replaced by scars, which appeared as defects on both the perfusion and metabolic images, and SCR was calculated from metabolic data, which is a percentage that indicates the degree of scarring. (iii) Perfusion-metabolic: MIS is a percentage that indicates the degree of perfusion metabolism mismatch, which refers to myocardial tissues show reduced or defective in MPI, while with normal or relatively increased 18 F-FDG uptake during MMI, and was calculated from combined perfusion and metabolic data.

PET Semi-quantitative Image Analysis
Semi-quantitative myocardial perfusion defects during rest were scored using the same 17-segment polar map; we used a five-point scale ranging from 0 to 4, corresponding to normal perfusion, slight reduction, moderate reduction, severe reduction, and no radiotracer uptake in each segment, respectively. 20,21 The summed rest scores in segments 1, 2, 7, 8, 13, 14, and 17; 5, 6, 11, 12, and 16; and 3, 4, 9, 10, and 15 were the regional SRS of LAD, LCx, and RCA, respectively. The possible value of SRS in LAD is 0-28, while it is 0-20 for LCx and RCA.

Coronary Angiography (CAG)
All patients underwent CAG by using the standard clinical technique within 1 month of PET/CT imaging. Experienced cardiologists visually interpreted the presence and degree of luminal stenosis of each coronary artery, and stenosis of a diameter of C 75% in at least one of the three major coronary arteries was considered as obstructive CAD. 15

Statistical Analysis
Statistical analysis was performed with the Statistical Program for Social Sciences (SPSS) software version 22.0 (SPSS, Chicago, IL, USA) 22 and the MedCalc software version 15.2.2 (MedCalc Software, Mariakerke, Belgium). 23 The reported statistical significance levels were all two sided, and P value \ 0.05 was considered indicative of statistically significant difference. Non-parametric rank-sum test or Chi-squared test (where appropriate) was used to compare differences in SRS, EXT, TPD, MBF, SCR, and MIS between the group with and without CAD. Correlation between each pair of quantitative indicators was assessed using Spearman's correlation coefficient (r). For a pair with |r| [ 0.8, the less significant indicator was eliminated. All possible combinations of the remained indicators were input into the classification models. We investigated seven types of ML algorithms, namely, logistic regression (LR), 24 linear discriminant analysis (LDA), 25 decision trees (DT), 26 support vector machines (SVM), 27 naive Bayes (NB), 28 K-nearest neighbors (KNN), 29 and random forest (RF). 30 Diagnostic performance was assessed using the sum of area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, and specificity, noted as sumScore. Statistically significant differences between AUCs were using DeLong's method. 31 Computer-generated random numbers were used to build a training set (159 vessels of 53 patients) and a validation set (105 vessels of 35 patients). The training set was used for indicator selection and model development, and the validation set was used for performance evaluation.

Subgroups Analysis
Four subgroup analyses were conducted on the validation set to verify whether the final selected model has good classification ability in specific populations. The four populations consisted of patients with the following: (i) old myocardial infarction (OMI) or/and revascularization history, (ii) with 0-1-or 2-3-vessel disease categorized based on the number of vessel with CAD, (iii) hypertension, and (iv) diabetes. Subgroup performance was compared with the SRS model by using ROC analysis.

Patient Characteristics
The patient characteristics are summarized in Table 1. As confirmed by CAG, 80 patients (91%) had obstructive CAD, and 8 patients (9%) had no obstructive CAD. Among the patients with obstructive CAD, 25 (31%) had single-vessel disease, 30 (38%) had twovessel disease, and 25 (31%) had three-vessel disease. Patients with obstructive CAD were mostly male and had higher diastolic and systolic blood pressure than those without obstructive CAD.

Selection of Predictors
From the analysis of 159 vessels in 53 patients on the training set, obstructive CAD in 92 (58%) vessels and non-obstructive CAD in 67 vessels (42%) were observed. The results for semi-quantitative SRS and five quantitative indicators (EXT, TPD, MBF, SCR, and MIS) derived from myocardial perfusion and metabolic imaging according to the presence of obstructive CAD in per-vessel analysis are shown in Table 2. Vessels with obstructive CAD showed significantly higher SRS, EXT, TPD, SCR, and MIS but lower MBF compared with the vessels without CAD (all P \ 0.001).

Model Performance
The sumScore (the sum of AUC, accuracy, sensitivity, and specificity) of all models under seven different ML methods is shown in Figure 1A. All 15 quantitative models showed higher sumScore than the semi-quantitative model (SRS) for each ML method (2.652-3.261 vs 2.412-2.703), except for model_3 (SCR) with a sumScore of 2.138-2.507 and model_123 in DT with a sumScore of 2.692. Details on the AUC, accuracy, sensitivity, and specificity of all models under the seven ML methods can be found in Supplementary  Table S2. Among the four individual indicators, perfusion and metabolic indicator MIS (model_4) achieved better performance than the remaining three and showed a sumScore of 2.816-3.042 vs 2.138-2.908. The metabolic indicator SCR showed the lowest sumScore of 2.138-2.507.
In the multivariate analysis, we separately selected the optimal model for each method on basis of the sumScore value as shown in Figure 1B. The optimal models for the DT, NB, KNN, and RF methods were model_134, model_12, model_24, and model_124, with sumScores of 3.042, 3.195, 3.154, and 3.147, respectively. The remaining LR, LDA, and SVM methods achieved the best performance in model_234, with sumScores of 3.229, 3.242, and 3.261, respectively. The lowest and highest sumScores of each model and corresponding ML methods as shown in Figure 1C and Supplementary Table S3. None of the highest sumScores of these 15 models was achieved by LR, while there are 6/15 models were achieved by SVM, we also noted that DT, RF, LDA, KNN, and NB also showed best performance for several models, while LR, KNN, DT, and RF showed lowest sumScore in 1, 3, 4, and 8 models, respectively, and none of the lowest sumScores were achieved by LDA, NB, and SVM.
The diagnostic performances of the seven optimal models in the validation set are shown in Figure 2 and Supplementary Table S4. Among these seven optimal models, SVM achieved the highest AUC (0.856), accuracy (0.810), and sensitivity (0.838) and moderate specificity (0.757) and thus was selected as the final optimal method, which includes the perfusion indicator MBF, the metabolic indicator SCR, and the joint indicator MIS, noted as model_234. The diagnostic performance of the semi-quantitative SRS model and 15 quantitative models on the validation set by using SVM is shown in Figure 3. The AUC, accuracy, sensitivity, and specificity of the semi-quantitative SRS model were 0.714, 0.657, 0.956, and 0.189, respectively.

Comparison with Semi-quantitative SRS
In the SVM method, the models including only perfusion indicators were model_1, model_2, and especially model_12, which had the best classification performance (sumScore: 3.192 vs 2.798-2.815). The model including only metabolic indicator was model_3 with a sumScore of 2.37. The 11 remaining models were combined perfusion and metabolic models, of which model_234 exhibited the best performance. Thus, model_12, model_3, and model_234 were selected as the representative models of perfusion, metabolic, and perfusion ? metabolic models, respectively. The ROCs of these three quantitative models and semi-quantitative SRS model are shown in Figure 4. The AUC of model_234 was higher than that of model_12 (0.856 vs 0.808) without significant difference (P = 0.084). Model_234 and model_12 both had significant higher AUC than SRS, while model_234 had stronger significant difference (P = 0.0008 vs 0.0127). Model_3 had lower AUC than SRS (0.651 vs 0.714). The quantitative

ROC Analysis in a Subgroups of Patients
As shown in Figure 5A, model_234 achieved significantly higher AUC than the SRS model (0.897 vs 0.775, P = 0.041) in detecting stenosis in 15 patients with OMI or/and revascularization history in the validation set.  At the cut-off C75% coronary stenosis, we divided the 35 validation patients into two groups of having 0-1or 2-3-vessel disease. For the 11 patients with 0-1-vessel disease ( Figure 5B), model_234 and model_SRS showed AUCs of 0.935 and 0.924, respectively, without significant difference. For the 24 patients with 2-3vessel disease ( Figure 5C), model_234 showed significantly higher AUC than model_SRS (0.839 vs 0.606, P = 0.005). This result indicates that model_234 consistently obtained good performance irrespective of the number of vessels with CAD. SRS showed limited ability in patients with multi-vessel disease.

DISCUSSION
Nuclear cardiology experts traditionally rely on visual evaluation and semi-quantitative analysis to interpret PET/CT MPI and MMI. 32,33 However, the diagnostic accuracy of these methods is limited and influenced by expert subjectivity. MPI quantitative analysis has been widely studied, and quantified PET MPI has substantially improved the accuracy of CAD diagnosis. 14,34 In China, the clinical routine for patients starts with MMI PET, followed by rest MPI PET. Therefore, we evaluated the joint value of PET rest MPI and MMI quantification in predicting obstructive CAD. Semi-quantitative SRS and five quantitative indicators (EXT, TPD, MBF, SCR, and MIS) were combined with seven ML algorithms to derive the optimal combination model and classification method.
Experimental results, which include MBF, SCR, and MIS, showed that model_234 based on the SVM method revealed the best diagnostic performance on the validation set, achieved the highest AUC and sumScore, and showed superiority over specific groups. As shown in Figure 5, model_234 achieved significantly higher AUC compared with model_SRS in patients with OMI or/and revascularization history, multi-vessel disease,  hypertension, and diabetes. This result indicates that model_234 has good diagnostic accuracy for patients with CAD and related chronic diseases.
In this study, luminal stenosis with a diameter of C 75% was defined as the disease according to the results of CAG. To assess whether 75% is a suitable classification cut-off, luminal stenosis with diameters C 50% 15,35 and C 90% 36 were defined on the same training and validation set, respectively. Thus, the ratio of normal to narrow in 264 vessels of 88 patients was 53:211 and 130:134, respectively. The ROCs of the best performing SVM models and semiquantitative SRS were compared for the identification of coronary artery stenosis, and the results are shown in Figure 6. At C 75% cut-off ( Figure 6B), the AUC values of the models obtained by C 50% ( Figure 6A) and C 90% ( Figure 6C) were reduced (0.622 and 0.755 vs 0.856). No significant difference was observed with the SRS model. Therefore, 75% is an appropriate classification cut-off to aid clinicians to accurately and effectively screen moderate and severe stenotic vessels requiring intervention, such as percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG).
Although CAG has been the gold standard for CAD diagnosis, [37][38][39][40] it has evident drawbacks, including its invasiveness and contrast agents, which can lead to complications and allergic reactions. Non-invasive imaging modalities, such as echocardiography and SPECT, are relatively insensitive for CAD detection. 41,42 Cardiac imaging with PET/CT is an accurate non-invasive and practical approach, and stress PET MPI is the most effective method to assess coronary microvascular disease. 43 Stress PET has been used in clinical practice for patients with mild CAD. [44][45][46][47] However, this method cannot be used for patients with moderate to severe CAD because it may cause acute cardiac events (ACE), such as shock. The cohort of this study mostly included patients with moderate to severe CAD and was only subjected to rest PET MMI and MPI scanning. This study investigated the diagnostic value of PET rest MPI and MMI for vascular stenosis in patients with obstructive CAD, and it indeed showed improved accuracy for the diagnosis of obstructive CAD. We believe that PET rest MPI and MMI can be used as the ''gatekeeper'' of CAG to reduce unnecessary CAG and save medical expenses for those patients who underwent both PET rest MPI and MMI, as this approach is currently not the standard clinical practice.
In the SVM method, we tried three kinds of kernel functions, namely linear-, radial basis-(RBF), and polynomial functions. The results of the three functions in 15 quantitative models are shown in Supplementary Figure S1. In most models (10/15), the linear function achieved the highest sumScores. Therefore, we finally chose linear as the kernel function of SVM in this work, and set the kernel scale to auto. When comparing the results of linear SVM and LR in 15 models, we can see that in most models (14/15), the sumScore of linear SVM is slightly higher or equal to LR. Detailed sumScore of each model is listed in Supplementary  Table S5. Both linear SVM and LR are linear classifiers, but on small-scale data sets, linear SVM is slightly better than LR, probably because linear SVM only considers a small part of the data that are most relevant to classification (i.e., support vectors), which improves the generalization and robustness; while LR considers the entire data set, each data point will affect the outcome of the decision.
This study showed some limitations. First, conventional stress PET MPI is the standard approach for patients to diagnose ischemia and CAD. Given the limitation of data, the experiment did not include stress study. Thus, the superiority of perfusion and metabolic combined approach over stress study was not assessed. We will continue to improve the experiments and strive to explore the value of joint FDG and perfusion. Second, the sample size was limited. Many patients undergo 13 N-NH 3 PET rest MPI, 18 F-FDG PET MMI, and CAG alone, but only a few underwent the three examinations in the short period of this study. Furthermore, patients with specific diseases (severe valvular heart disease, severe myocarditis, and arrhythmia) and incomplete clinical information were excluded. Thus, the study population is difficult to expand in a short period. Current data clearly showed significant differences between PET cardiac imaging quantification and semiquantification. Although the findings were sufficient to confirm the diagnostic value of joint PET rest MPI and MMI quantification for obstructive CAD diagnosis, a substantial cohort or multiple centers is required to validate the present model in future studies.
To conclude, this study investigated the diagnostic value of joint PET rest MPI and MMI quantitative indicator combined with ML for obstructive CAD diagnosis. The multivariate model (MBF, SCR, and MIS) combined with SVM outperformed other methods and thus may aid clinicians in accurately predicting the presence of obstructive CAD without performing invasive CAG. But in this work, we only investigated the superiority of the FDG/resting perfusion approach over semi-quantitative SRS. We have not investigated, and thus not shown, the superiority of joint FDG and resting perfusion approach over stress MPI.

NEW KNOWLEDGE GAINED
The diagnostic performance of model_234 (MBF, SCR, and MIS) with SVM for coronary stenosis detection was better than that of other machine learning methods. The AUC of quantitative model_234 combining MPI and MMI information was significantly higher than that of the semi-quantitative model_SRS whether in the validation set or four specific subgroups.