Radiomics-based precision phenotyping identifies unstable coronary plaques from computed tomography angiography

Abstract

Background:

It remains debated whether any single coronary atherosclerotic plaque within the vulnerable patient exhibits unique morphology conferring an increased risk of clinical events.

Objectives:

To precisely phenotype culprit and nonculprit lesions in myocardial infarction (MI) and lesions in stable coronary artery disease (CAD) using coronary computed tomography angiography (CCTA)-based radiomic analysis.

Methods:

Sixty patients with acute MI prospectively underwent CCTA prior to invasive angiography and were matched to 60 patients with stable CAD. For all coronary lesions, high-risk plaque (HRP) characteristics were qualitatively assessed, followed by semi-automated plaque quantification and extraction of 1,103 radiomic features. Machine learning models were built to assess the additive value of radiomic features for discriminating culprit lesions over and above HRP and plaque volumes.

Results:

Culprit lesions had higher mean volumes of noncalcified plaque (NCP) and low-density noncalcified plaque (LDNCP) compared with the highest-grade stenosis nonculprits and highest-grade stenosis stable CAD lesions (NCP: 138.1mm³ vs. 110.7mm³ vs. 102.7mm³; LDNCP: 14.2mm³ vs. 9.8mm³ vs. 8.4mm³; both p_trend<0.01). In multivariable linear regression adjusted for NCP and LDNCP volumes, 14.9% (164/1,103) of radiomic features were associated with culprits and 9.7% (107/1,103) were associated with the highest-grade stenosis nonculprits (critical p<0.0007) when compared with highest-grade stenosis stable CAD lesions as reference. Hierarchical clustering of significant radiomic features identified 9 unique data clusters (latent phenotypes): 5 contained radiomic features specific to culprits, 1 contained features specific to highest-grade stenosis nonculprits, and 3 contained features associated with either lesion type. Radiomic features provided incremental value for discriminating culprit lesions when added to a machine learning model containing HRP and plaque volumes (area under the receiver operating characteristic curve 0.86 vs. 0.76, p=0.004).

Conclusions:

Culprit lesions and highest-grade stenosis nonculprit lesions in MI have distinct radiomic signatures. Within the vulnerable patient may exist individual vulnerable plaques identifiable by CCTA-based precision phenotyping.

INTRODUCTION

The concept of the “vulnerable patient” is well-established, with coronary atherosclerotic disease burden being a strong predictor of subsequent acute coronary syndrome (ACS) in individuals with stable coronary artery disease (CAD) (1). However, it is debated whether any single “vulnerable plaque” in these patients exhibits unique morphological features conferring an increased risk of clinical events over total atheroma burden (1,2). Intracoronary imaging studies of culprit lesions at the time of ACS have identified vulnerable plaque features such as a large lipid-rich necrotic core, thin fibrous cap, and microcalcifications (3). Yet, invasive plaque characterization remains technically difficult and yields a low positive predictive value for future ACS (4). Coronary computed tomography angiography (CCTA) is a noninvasive first-line modality for the evaluation of suspected CAD (5). Qualitative plaque characteristics and quantitative attenuation-based plaque measures on CCTA have demonstrated prognostic value for patient-level risk of coronary events (6–8). Beyond these conventional parameters, however, may exist latent morphological plaque features which are not captured by current CCTA analysis techniques. Radiomics, by extracting thousands of computational quantitative features from medical images, enables precision phenotyping of diseases in vivo (9). Prior reports have shown CCTA-based radiomic analysis to reliably detect vulnerable plaques in patients with stable CAD when compared to invasive and histological reference standards (10,11). This study aimed to evaluate whether culprit lesions in acute myocardial infarction (MI) have a distinct radiomic signature on CCTA compared with nonculprit lesions in the same patients and with lesions in stable CAD patients. We also sought to determine the incremental value of radiomic features for discriminating culprit lesions beyond current state-of-the-art plaque assessment in CCTA.

METHODS

Study Population

This was a post-hoc analysis of the prospective INFLAME study (Inflammation of pericoroNary Fat and its association with coronary atheroscLerosis Assessed by coMputEd tomography coronary angiography; Australian New Zealand Clinical Trials Registration ACTRN12618001058268). The patient selection and study design have been previously published (12). Briefly, 60 consecutive patients presenting with acute MI at MonashHeart, Melbourne, Australia were recruited to undergo CCTA within 48 hours of admission, prior to invasive coronary angiography (ICA). These patients were matched to 60 outpatients with stable CAD (defined by stable exertional symptoms and/or inducible myocardial ischemia on stress testing) who underwent CCTA during the same time period. Matching was performed for age, sex, body mass index (BMI), risk factors (diabetes, hypertension, dyslipidemia, smoking), medications, and CT tube voltage (Supplemental Appendix). The study was approved by the local Human Research Ethics Committee (Monash Health HREC 15244A) and all MI patients provided written informed consent.

MI and Culprit Lesion Adjudication

Adjudication of MI was performed by a panel of 2 cardiologists (A.L. and D.W.) using the Fourth Universal Definition (13). Cases were classified as ST-segment elevation MI (STEMI) or non-ST-segment elevation MI (NSTEMI). One culprit lesion per patient was identified at the time of ICA by an experienced interventional cardiologist (the primary operator) blinded to CCTA findings, using a combination of angiographic appearance (stenosis severity and lesion morphology), electrocardiogram findings, and wall motion abnormalities on left ventriculography or echocardiography. Independent adjudication of culprit lesions was performed post-hoc by a second interventional cardiologist (D.W.) blinded to the decisions of the primary operator. Ambiguous cases or cases with multiple candidate culprits were resolved by consensus. All other lesions were deemed nonculprit lesions.

Ascertainment of Risk Factors

Pre-existing cardiovascular risk factors and medications were documented on admission for the MI cohort and ascertained by review of electronic medical records for the stable CAD cohort. Risk factors are defined in the Supplemental Appendix.

CCTA Acquisition

All scans were performed on a 320-detector-row CT scanner (Aquilion ONE ViSION, Canon Medical Systems, Otawara, Japan) as previously described (14) (Supplemental Appendix).

Qualitative Plaque Analysis

CCTA analysis was performed at the Cedars-Sinai Medical Center core laboratory by a level III-experienced reader (A.L.) blinded to clinical data and ICA findings. Using axial and multiplanar reconstruction views, all coronary segments ≥2 mm were evaluated according to an 18-segment model (15). For each patient, we calculated the segment involvement score and segment stenosis score (16) as qualitative measures of the extent and severity of CAD. Each coronary lesion was assessed for adverse plaque characteristics (6,17) (APCs; Supplemental Appendix). High-risk plaque (HRP) was defined by the presence of ≥2 APCs including positive remodeling, low attenuation plaque, spotty calcification, or napkin-ring sign. A second expert reader (D.H.) in the core laboratory independently assessed all lesions for the presence of HRP; interobserver agreement was moderate (κ=0.58).

Quantitative Plaque Analysis

Standardized plaque measurements were performed using semiautomated research software (Autoplaque v2.5, Cedars-Sinai Medical Center, Los Angeles, CA, USA) (18,19) (Supplemental Appendix). Plaque volume (mm³) was calculated on a per-lesion level for the following plaque components: total plaque (TP), calcified plaque (CP), noncalcified plaque (NCP), and low-density noncalcified plaque (LDNCP; defined by an attenuation of <30 Hounsfield units [HU]). The respective plaque burdens (%) were calculated as: plaque volume / analyzed vessel volume × 100 (20). Plaque volume and burden measurements for the entire coronary tree were summarized on a per-patient level. Per-lesion plaque composition (%) by CP, NCP, and LDNCP components was calculated as: plaque component volume / total plaque volume × 100 (19). Additionally, for each lesion, the software automatically quantified plaque length, percent diameter stenosis, and vessel remodeling index (Supplemental Appendix). Stenosis severity was further categorized according to the Coronary Artery Disease Reporting and Data System (CAD-RADS): 0 (0%), 1 (1–24%), 2 (25–49%), 3 (50–69%), 4 (70–99%), 5 (100%) (21). Following CCTA plaque analysis, ICA-determined culprit lesions were co-registered to lesions on CCTA (Supplemental Appendix).

Radiomic Plaque Analysis

All voxels defined as plaque by the semiautomated software (Autoplaque v2.5) were extracted as DICOM images and loaded into the open-source Radiomics Image Analysis (RIA v1.4.2) software package in the R environment (22). This process was performed for: 1) culprit lesions; 2) all nonculprit lesions as one segmentation; 3) all lesions in stable CAD as one segmentation; 4) highest-grade stenosis non-culprit lesions; 5) highest-grade stenosis lesions in stable CAD. The images were discretized to 8, 16, and 32 equally sized bins with identical HU ranges, creating 3 replicas of each image. A total of 1,103 radiomic features were calculated for each segmentation, including: 44 first-order metrics describing the distribution of HU values; 342 gray-level co-occurrence matrix (GLCM) textural parameters describing how often voxels of a similar value occur next to each other; 33 gray-level run-length matrix (GLRLM) textural parameters describing how often a given number of similar-value voxels occur next to each other; and 684 geometric parameters describing spatial characteristics such as shape, size, or volume. A detailed description of each radiomic feature has been previously published (23–25). Central Illustration shows the workflow for plaque segmentation and radiomic analysis.

Central Illustration. Radiomics-based precision phenotyping in myocardial infarction.

Plaque segmentation was performed on CCTA images (case example of a culprit lesion shown) using semiautomated software. Noncalcified plaque is displayed in red, low-density noncalcified plaque in orange, and vessel lumen in blue. Radiomic features were extracted from plaque-containing voxels and analyzed using hierarchical clustering. Nested machine learning models were built using conventional plaque parameters and cluster-derived radiomic features; the performance of these models in discriminating culprit lesions is shown on the right. AUC = area under the receiver operating characteristic curve; HRP = high-risk plaque.

Machine Learning

Model Building and Training

Given the high dimensionality of our dataset and the complex inter-relationships that exist between plaque features, we applied machine learning (ML) to our radiomic analysis; an approach previously shown to be useful for identifying advanced atherosclerotic lesions (10). To examine the incremental value of radiomic features beyond conventional plaque analysis for discriminating culprit lesions from the highest-grade stenosis nonculprit and stable CAD lesions, we built 3 nested ML models using CCTA-based plaque parameters: model 1: presence of HRP; model 2: model 1 plus total plaque volume plus volumes and compositions of NCP and LDNCP; and model 3: model 2 plus 18 radiomic features with the smallest p-values from the 9 distinct data clusters (see Statistical Analysis): 9 features associated with culprits and 9 features associated with the highest-grade stenosis nonculprits. To examine the incremental value of radiomic features for culprit lesion discrimination beyond stenosis, we performed an additional ML analysis with the following parameters: model 1: stenosis severity by CAD-RADS category; model 2: model 1 plus the 18 radiomics features. We employed supervised ML with XGBoost, a state-of-the-art boosted ensemble algorithm which has demonstrated high performance within the domains of computer science and medicine (26). XGBoost uses an ensemble of gradient-boosted decision trees, combining multiple weak classifiers (one-level decision trees) to produce a single strong classifier, which can improve prediction modeling. We trained and tested all 3 of our ML models for a binary classification task (culprit lesion vs. other lesion types) using 10 times repeated 10-fold cross-validation (27) (Supplemental Appendix). All ML analyses were performed using Python 3.8.5, scikit-learn 0.23.2, and XGBoost 1.3.3.

External Validation

We evaluated the performance of the radiomics-based ML model in an external validation cohort from University Hospital Erlangen, Erlangen, Germany. This consisted of 19 patients admitted with ACS who underwent CCTA prior to ICA, as well as 16 matched controls with stable CAD (28). The patient selection and CCTA imaging protocol are detailed in the Supplemental Appendix. Clinical and CCTA characteristics of this cohort have been previously published (28). Qualitative, quantitative, and radiomic plaque analysis were performed using the same methodology described in the present study.

Statistical Analysis

At the patient level, MI patients were compared 1:1 to matched patients with stable CAD. At the lesion level, culprit lesions were compared to: 1) the average of all nonculprit lesions (within-subject) and the average of all stable CAD lesions (between-subject); and 2) the highest-grade stenosis nonculprit lesion (within-subject) and the highest-grade stenosis stable CAD lesion (between-subject), as determined by quantitative plaque analysis. We specifically examined highest-grade stenosis lesions given that coronary lumen stenosis remains the dominant measure of CAD severity used in clinical practice to guide patient management.

Continuous variables are presented as mean ± standard deviation or median (interquartile range), as appropriate. Categorical variables are expressed as frequencies (%). Comparisons of continuous variables were performed using an independent samples t-test or a one-way ANOVA with Tukey’s post-hoc test. Categorical variables were compared with a Fisher’s exact or Chi-square test. Interobserver agreement was calculated using the κ statistic. For clinical and conventional plaque parameters, a 2-sided p-value <0.05 indicated statistical significance.

For lesion-level radiomic characterization, principal component analysis identified 72 unique radiomic features which accounted for 99.5% of the variance in the data (29). Hence, applying Bonferroni correction, p-values <0.0007 (0.05/72) were considered statistically significant. To examine the association of culprit and nonculprit lesions with each of the 1,103 radiomic features, we performed linear regression analysis using lesion types as categorical predictors with stable CAD lesions as reference. We then adjusted for volumes of CP, NCP, and LDNCP to correct for any potential intrinsic correlation between plaque volume and morphology. From the multivariable linear regression, we identified the significant radiomic features associated with either culprit or nonculprit lesions and performed pairwise linear regression between these features. The resultant intra-pair R² values were used as distance measures for hierarchical clustering. The elbow method based on the total within-cluster sum of squares was used to calculate the optimal number of distinct data clusters (latent phenotypes): 15 for culprit vs. all nonculprits; 9 for culprit vs. highest-grade stenosis nonculprit.

For clinical intuition, we performed conventional statistical logistic regression on the same models that were used in the ML analysis: model 1: HRP; model 2: model 1 plus quantitative plaque parameters; and model 3: model 2 plus radiomic features. All 3 logistic regression models were 10-fold cross-validated with the same data folds used in ML. To avoid including highly correlated radiomic features in statistical model 3, we performed backward stepwise logistic regression at a Wald p-value of 0.10, resulting in 8 radiomic features being retained in the final model. Receiver-operating characteristic analysis was used to evaluate the discriminatory ability of all models for identifying culprit lesions, and the area under the curve (AUC) values were compared using the method of Delong et al. (30). All statistical analyses were performed using R 3.5, R Studio 1.2.1335, and Stata 14.0 (StataCorp, College Station, TX, USA).

RESULTS

Baseline Patient Characteristics

Table 1 summarizes characteristics of the study population according to clinical presentation. The 2 cohorts were well-matched for age, sex, BMI, risk factors, medications, and CT tube voltage. Of patients with MI, 55 (91.7%) presented with NSTEMI and 5 (8.3%) with post-thrombolysis STEMI. All patients had stenting of the culprit lesion and 9 (15.0%) patients had stenting of the highest-grade stenosis nonculprit lesion during the index admission.

Table 1.

Patient characteristics according to clinical presentation

	Acute MI (n = 60)	Stable CAD (n = 60)	p value
Clinical characteristics
Age, yrs	59.9±11.6	60.2±11.3	0.68
Body mass index, kg/m2	28.3±5.9	28.7±5.0	0.29
Male sex	52 (86.7)	52 (86.7)	1.00
Hypertension	44 (73.3)	41 (68.3)	0.49
Diabetes	15 (25.0)	13 (21.7)	0.53
Dyslipidemia	32 (53.3)	35 (58.3)	0.36
Smoking	24 (40.0)	20 (33.3)	0.21
Family history of CAD	25 (41.7)	27 (45.0)	0.47
Baseline medications
Antiplatelet	11 (18.3)	10 (16.7)	0.43
Statin	15 (25.0)	17 (28.3)	0.17
Beta-blocker	10 (16.7)	12 (20.0)	0.38
ACE-I or ARB	14 (23.3)	14 (23.3)	0.26
CCTA acquisition parameters
Heart rate, bpm	53.9±5.3	54.8±8.4	0.58
Tube voltage			1.00
100 kV	22 (36.7)	22 (36.7)
120 kV	38 (63.3)	38 (63.3)
Contrast dose, mL	77.6±11.0	76.2±13.9	0.83
Radiation dose, DLP	284.8±156.9	246.1±171.4	0.26

Values are expressed as n (%) or mean ± standard deviation.

Bold p values indicate statistical significance.

ACE-I = angiotensin converting enzyme inhibitor; ARB = angiotensin receptor blocker; CAD = coronary artery disease; CCTA = coronary computed tomography angiography; DLP = dose length product; HDL = high-density lipoprotein; LDL = low-density lipoprotein; MI = myocardial infarction.

Qualitative Plaque Parameters

Patients with MI had a higher average number of lesions compared to patients with stable CAD (5.6±3.9 vs. 3.9±2.7, p=0.03). HRP was more frequent in MI patients than in stable CAD patients (70.0% vs. 31.7%, p<0.001), as was each component of HRP (all p<0.05; Table 2).

Table 2.

Per-patient level plaque analysis

	Acute MI (n=60)	Stable CAD (n=60)	p value
Total number of lesions	5.6±3.9	3.9±2.7	0.03
Qualitative parameters
Segment involvement score	5.8±3.1	5.1±2.5	0.38
Segment stenosis score	14.2±7.9	9.7±5.2	0.005
High-risk plaque present	42 (70.0)	19 (31.7)	<0.001
Positive remodeling present	50 (83.3)	33 (55.0)	<0.001
Low attenuation plaque present	46 (76.7)	14 (23.3)	<0.001
Spotty calcification present	20 (33.3)	10 (16.7)	<0.001
Napkin-ring sign present	16 (26.7)	4 (6.7)	<0.001
Quantitative parameters
Plaque volume, mm3
Total plaque	596.9±408.1	278.9±289.0	<0.001
Calcified plaque	75.1±98.5	29.6±45.5	0.002
Noncalcified plaque	521.9±332.2	229.3±248.1	<0.001
Low-density noncalcified plaque	36.8±24.5	17.2±21.1	<0.001
Plaque burden, %
Total plaque	45.2±7.1	40.3±8.1	0.001
Calcified plaque	4.3±3.5	3.4±2.9	0.14
Noncalcified plaque	41.0±8.5	37.0±9.7	0.02
Low-density noncalcified plaque	3.1±1.6	2.7±1.5	0.03

Values are expressed as n (%) or mean ± standard deviation.

Bold p values indicate statistical significance.

CAD = coronary artery disease, MI = myocardial infarction.

Culprit lesions more often exhibited each of the 4 APCs compared to all nonculprit lesions in MI and all lesions in stable CAD (all p<0.05; Supplemental Table 1). When compared to the highest-grade stenosis nonculprit and stable CAD lesions, culprit lesions were more likely to exhibit low attenuation plaque or positive remodeling (all p<0.05; Table 3). There were no significant differences in the prevalence of APCs or HRP between nonculprit lesions and stable CAD lesions (Supplemental Table 1 and Table 3).

Table 3.

Per-lesion level plaque analysis for culprit vs. highest-grade stenosis nonculprit in MI vs highest-grade stenosis lesion in stable CAD

	Culprit lesion (n=60)	Highest stenosis nonculprit lesion (n=60)	Highest stenosis stable CAD lesion (n=60)	p value	p values for post-hoc comparisons
					Culprit vs. nonculprit	Culprit vs. stable	Nonculprit vs. stable
Qualitative parameters	n (%)	n (%)	n (%)
High-risk plaque present	32 (53.3)	15 (25.0)	10 (16.7)	<0.001	0.005	<0.001	0.86
Positive remodeling present	40 (66.7)	28 (46.7)	20 (33.3)	0.001	0.04	0.001	0.58
Low attenuation plaque present	42 (70.0)	9 (15.0)	9 (15.0)	<0.001	<0.001	<0.001	1.00
Spotty calcification present	10 (16.7)	10 (16.7)	6 (10.0)	0.09	1.00	0.86	0.86
Napkin-ring sign present	10 (16.7)	7 (11.7)	4 (6.7)	0.53	1.00	0.72	1.00
Quantitative parameters
Plaque volume, mm3
Total plaque	150.6±86.8	125.3±129.6	116.3±75.4	<0.001	<0.001	<0.001	0.09
Calcified plaque	12.5±23.4	15.7±28.3	16.5±22.1	0.14	0.21	0.13	0.83
Noncalcified plaque	138.1±81.9	110.7±102.4	102.7±63.2	0.004	0.005	0.03	0.35
Low-density noncalcified plaque	14.2±10.7	9.8±8.6	8.4±7.9	0.001	0.008	0.001	0.42
Plaque burden, %
Total plaque	60.2±10.9	50.1±11.6	49.9±10.2	<0.001	<0.001	<0.001	0.57
Calcified plaque	3.8±5.6	4.2±5.1	4.4±3.7	0.21	0.92	0.28	0.14
Noncalcified plaque	56.4±13.7	46.5±12.7	42.3±11.6	<0.001	<0.001	<0.001	0.16
Low-density noncalcified plaque	6.7±5.5	4.2±2.7	3.1±2.2	<0.001	0.008	<0.001	0.03
Plaque composition, %
Calcified plaque	7.3±10.7	9.3±11.4	9.6±10.5	0.14	0.87	0.70	0.94
Noncalcified plaque	92.7±10.7	90.7±11.4	90.4±10.5
Low-density noncalcified plaque	10.1±7.8	7.0±4.7	6.9±4.1	0.001	0.006	0.004	0.96
Plaque length, mm	17.3±7.6	19.7±13.7	18.5±11.1	0.11	0.19	0.76	0.85
Diameter stenosis, %	86.9±12.6	58.1±22.3	55.4±20.9	<0.001	<0.001	<0.001	0.20
Remodeling index	1.38±0.3	1.25±0.2	1.21±0.2	0.17	0.32	0.15	0.74

Values are expressed as n (%) or mean ± standard deviation. Bold p values indicate statistical significance.

CAD = coronary artery disease, MI = myocardial infarction.

Quantitative Plaque Parameters

Patients with MI had greater volumes of all plaque components compared to patients with stable CAD (all p<0.05; Table 2). Burdens of TP (45.2±7.1% vs. 40.3±8.1%, p=0.001), NCP (41.0±8.5% vs. 37.0±9.7%, p=0.02), and LDNCP (3.1±1.6% vs. 2.7±1.75, p=0.03) were higher in MI versus stable CAD patients; with no significant differences in CP burden (p=0.14).

Culprit lesions had a mean TP volume of 150.6±86.8 mm³, comprising 12.5±23.4 mm³, 138.1±81.9 mm³, and 14.2±10.7 mm³ of CP, NCP, LDNCP, respectively. The volumes and burdens of TP, NCP, and LDNCP were higher in culprits compared to nonculprits and stable CAD lesions (all p<0.05; Supplemental Table 1 and Table 3); no significant differences were observed for CP (all p>0.05). LDNCP composition was greater in culprits (10.1±7.8%) compared to all nonculprits and all stable CAD lesions (5.9±3.3% and 5.8±2.7%, respectively; both p<0.05); and compared to the highest-grade stenosis nonculprits and highest-grade stable CAD lesions (7.0±4.7 and 6.9±4.1, respectively; both p<0.05). LDNCP burden was greater in the highest-grade nonculprits than in the highest-grade stable CAD lesions (4.2±2.7% vs. 3.1±2.2%, p=0.03); with no significant differences in the remaining quantitative parameters between the 2 plaque types (all p>0.05).

Radiomic Analysis

In linear regression analysis, 349 (31.6%) of 1,103 radiomic features were significantly associated with culprit lesions and 273 (24.8%) were associated with all nonculprit lesions, when compared with all stable CAD lesions as reference. Following multivariable adjustment for volumes of CP, NCP, and LDNCP, 367 (33.3%) radiomic features were associated with culprits, whereas no features were associated with nonculprits (Supplemental Figure 1).

Linear regression performed using culprits and the highest-grade stenosis nonculprits showed an association with 186 (16.9%) and 47 (4.3%) radiomic features, respectively, referenced against the highest-grade stenosis stable CAD lesions. When adjusted for plaque volumes, 164 (14.9%) and 107 (9.7%) features, respectively, remained statistically significant (Figure 1). A greater proportion of textural (GLCM and GLRLM) and geometric features, rather than first-order metrics, differed between culprit and nonculprit lesions (Supplemental Table 2).

Figure 1. Manhattan plot of p-values for associations of culprit lesions and highest-grade stenosis nonculprit lesions with radiomic features.

P values are displayed on the y axis for each of the 1,103 radiomic parameters lined up on the x axis, color-coded by category. Upper panels show results from unadjusted linear regression analysis; bottom panels show results following adjustment for volumes of calcified, noncalcified, and low-density noncalcified plaque. Points above the red line (p=0.0007) indicate radiomic features significantly associated with either culprit or highest-grade stenosis nonculprit lesions. The highest-grade stenosis lesion in stable CAD was used as reference.

On cluster analysis of significant radiomic features, we identified 15 statistically unique data clusters (latent phenotypes); each cluster contained radiomic features associated exclusively with culprit lesions (Supplemental Figure 2). When comparing culprits to the highest-grade stenosis nonculprits, we identified 9 distinct clusters. Of these, 5 clusters contained only radiomic features specific to culprits, 1 contained only features specific to nonculprits, and 3 contained features associated with either lesion type (1 cluster contained 6 features common to both lesion types) (Figure 2).

Figure 2. Clustering dendrogram and heatmap of significant radiomic features associated with culprit vs. highest-grade stenosis nonculprit lesion.

There were 9 distinct clusters (red boxes) among the 265 significant radiomic features when comparing the culprit lesion and highest-grade stenosis nonculprit lesion (middle panel). Of these, 5 clusters contained only radiomic features specific to culprit lesions (pink), 1 cluster contained only features specific to nonculprit lesions (green), and 3 clusters contained features associated with either lesion type (1 cluster contained 6 features associated with both lesion types, as shown in blue). The highest-grade stenosis stable CAD lesion was used as reference.

Machine Learning for Identification of Culprit Lesions

In the ML models, adding quantitative plaque parameters to HRP resulted in a significantly higher AUC for discriminating culprit lesions (0.76, 95% confidence interval [CI] 0.68–0.83) compared to HRP alone (0.65, 95% CI 0.57–0.74; p=0.01). Further addition of radiomic features provided incremental discriminatory value over and above HRP and quantitative plaque parameters (AUC 0.86, 95% CI 0.80–0.92; p=0.004; Central Illustration). The ranking of feature importance for identifying culprit lesions in the final ML model is shown in Figure 3. Textural (GLCM) and geometric radiomic features occupied the top 7 positions overall, while LDNCP volume and composition were highest-ranked among quantitative plaque parameters. When evaluated in the external validation cohort, the radiomics-based ML model exhibited an AUC of 0.84 (95% CI 0.74–0.94) for culprit lesion discrimination.

Figure 3. Feature importance for machine learning identification of culprit lesions.

Ranking of feature importance in the final ML model incorporating HRP, quantitative plaque parameters, and radiomic features. The solid bars and error bars represent the mean feature importance value and standard deviation, respectively.

ML analysis in the main cohort also showed radiomic features alone (AUC 0.84, 95% CI 0.78–0.91) to outperform the combination of HRP and quantitative plaque parameters (p=0.009) and HRP alone (p<0.001) for discriminating culprit lesions. The addition of radiomic features to a ML model with stenosis severity provided incremental value for culprit lesion discrimination, with an AUC increase from 0.83 (95% CI 0.77–0.89) to 0.89, (95% CI 0.84–0.94; p=0.002).

Performance of Conventional Statistical Models

In logistic regression analysis, a model with radiomic features, plaque volumes, and HRP (AUC 0.80, 95% CI 0.74–0.85) outperformed a model with plaque volumes and HRP (AUC 0.71, 95% CI 0.65–0.76; p=0.005) and HRP alone (AUC 0.63, 95% CI 0.57–0.69; p<0.001) for culprit lesion discrimination.

DISCUSSION

Our primary findings are: 1) culprit lesions in MI have a distinct radiomic phenotype on CCTA compared with nonculprit lesions and stable CAD lesions; 2) among all nonculprits, the highest-grade stenosis lesion carries a unique radiomic signature; 3) the distinguishing radiomic features in both scenarios are textural and geometric; and 4) radiomic features provide incremental value for discriminating culprit lesions beyond current state-of-the-art CCTA-based plaque assessment.

Radiomic analysis enables precision phenotyping of diseases in vivo, by extracting thousands of computational quantitative features from an imaging region of interest which potentially reflect the underlying pathophysiology (9). Coronary plaques comprise several different histological components which vary in their CT HU attenuation values (31). This has enabled the noninvasive assessment of plaque morphology, with numerous CCTA-derived qualitative (6,17) and quantitative (7,8) parameters associating with an increased risk of subsequent ACS. However, visual plaque assessment is prone to inter-reader variability (32), and semiautomated plaque quantification relies on absolute voxel attenuation values (31) without considering the complex spatial relationship between voxels. Radiomics provides an objective, automated, data-driven description of plaque morphology using measures of heterogeneity or shape which are mostly invisible to the human eye (24). This approach has outperformed conventional qualitative and quantitative metrics for identifying vulnerable plaques in stable patients undergoing CCTA when compared to invasive reference standards (10). Further, in an ex vivo study, radiomic features improved the accuracy of CCTA for identifying histologically-verified thin-cap fibroatheroma (11). The present analysis extends these findings by demonstrating that radiomic analysis can reliably detect and characterize MI-causing unstable plaques at the time of event.

Several prior studies have utilized CCTA for plaque assessment in patients presenting with ACS. Early reports showed a higher prevalence of qualitative APCs in culprit lesions compared to lesions in stable CAD (33,34). Subsequent investigators used software applications to quantify plaque volumes and burdens in the setting of ACS (20,35). The current study is the first to perform radiomic phenotyping of plaque in patients with acute MI, and to compare radiomic signatures among culprit, nonculprit, and stable CAD lesions. By extracting the maximal quantitative information from routine CCTA images, radiomics can provide an insight into fine plaque structure and capture latent traits associated with plaque instability that are typically only explored on the molecular level. We showed textural and geometric features – rather than first-order HU-based metrics – to define the radiomic signature of culprit lesions, and to provide incremental discriminatory value beyond current qualitative and quantitative plaque measures. Textural features characterize plaque composition, expressed by the spatial relationship of voxels on CCTA, while geometric features capture the complex three-dimensional structure of plaques. Certainly, there is evidence from histological studies that the spatial distribution and localization of different plaque components influence plaque vulnerability (17,36). Furthermore, we have previously shown geometric features to associate with echo-attenuated plaque on intravascular ultrasound and thin-cap fibroatheroma on optical coherence tomography (10); potentially reflecting a large lipid pool. Hence, comprehensive CCTA-based plaque analysis integrating conventional parameters with radiomic features may improve our understanding of the morphological factors that act synergistically to increase the risk of ACS.

The “vulnerable patient” concept is well-established (1), and overall coronary atherosclerotic burden quantified by CCTA has shown predictive value for future MI independently of obstructive stenoses or APCs (8,37). Consistent with our prior work (20), the present analysis demonstrated higher volumes and burdens of TP, NCP, and LDNCP in patients with MI compared to patients with stable CAD. Similar results were observed at the lesion level when comparing culprits to nonculprits in MI and to lesions in stable CAD. On radiomic analysis of the entire coronary tree in both cohorts, only the culprit lesion was significantly associated with radiomic features following adjustment for plaque volumes. In single-lesion comparisons, the highest-grade stenosis nonculprit exhibited a distinct radiomic signature compared to the highest-grade stenosis stable CAD lesion. The distinguishing radiomic features were textural and geometric parameters – several of which were also present in culprit lesions – and may reflect vulnerable plaque morphology. Indeed, we showed LDNCP burden to be greater in the highest-grade stenosis nonculprit compared to the highest-grade stable CAD lesion. This is consistent with intracoronary imaging studies demonstrating obstructive nonculprits in MI to more frequently harbor thin-cap fibroatheroma compared to nonobstructive nonculprits (38) or obstructive lesions in stable CAD (39); and suggests that obstructive disease is indirectly associated with subsequent events due to its relationship with plaque burden and morphology. Importantly, recent evidence shows that stenting of nonculprit lesions in MI based on stenosis severity significantly reduces the risk of future MI or cardiac death (40). Overall, our noninvasive imaging findings support the hypothesis that within the vulnerable patient exists one or more vulnerable plaques with distinct morphologies, as evidenced by the radiomic signatures of culprit lesions and highest-grade stenosis nonculprit lesions.

Our study results reflect the morphologies of unstable or disrupted coronary plaques at the time of event, and are hypothesis-generating. It is reasonable to postulate that not-yet-disrupted vulnerable plaques would share similar radiomic features to culprit lesions. Indeed, prior studies have shown that qualitative morphological characteristics do not differ significantly between intact vulnerable plaques and ruptured plaques on CCTA (41,42). Given that established definitions of plaque vulnerability are based on event-causing lesions examined by histology or intracoronary imaging (3,36), the use of CCTA radiomics to phenotype plaque in MI advances the role of noninvasive imaging in the detection of rupture-prone lesions. Future studies should examine plaque radiomic features of stable patients in whom the precursor lesions associated with future ACS have been identified. Moreover, the ability to modify the radiomic signature of the highest-grade stenosis nonculprit lesion in patients with MI through lipid-lowering or anti-inflammatory therapies should be explored. Notably, PROSPECT-ABSORB (Providing Regional Observations to Study Predictors of Events in the Coronary Tree II Combined with a Randomized, Controlled, Intervention Trial) recently showed pre-emptive stenting of nonculprit vulnerable plaques in MI based on intracoronary imaging to be associated with favorable long-term outcomes (43). This paves the way for larger randomized interventional trials, wherein CCTA could be used to help identify candidate vulnerable plaques. Finally, while plaque quantification and radiomic analysis remain investigational, rapid advancements in hardware, software, and artificial intelligence algorithms will facilitate the future implementation of these techniques into clinical practice. In the present study, radiomic extraction followed semi-automated plaque segmentation. The ultimate goal is for fully automated, artificial intelligence-based plaque and radiomic analysis to be embedded into routine CCTA reporting software, providing real-time measurements and flagging vulnerable plaques for clinician review.

Limitations

This study has several limitations. The case-control design may limit the generalizability of our findings to larger, unselected populations. We utilized a small external validation cohort, and our radiomics-based ML model needs further validation in datasets from different centers and CT scanners. There was a predominance of NSTEMI cases, requiring strict adjudication of culprit lesions using invasive angiographic, ECG, and echocardiographic findings. Radiomic analysis was performed on culprit lesions early following MI, and thus the some of the identified radiomic features may reflect post-ACS plaque inflammation. We do not have correlation with intracoronary imaging or histological specimens for the current plaque radiomic signatures, however our prior work has shown radiomic features on CCTA to accurately identify vulnerable plaques as defined by these reference standards in stable patients (10,11).

CONCLUSION

Culprit lesions and highest-grade stenosis nonculprit lesions in MI have distinct radiomic signatures compared with lesions in stable CAD. Within the vulnerable patient, there may exist individual vulnerable plaques which are identifiable using CCTA-based precision phenotyping.

PERSPECTIVES.

Competency in Medical Knowledge

Culprit lesions and highest-grade stenosis nonculprit lesions in MI have distinct radiomic signatures on CCTA compared with lesions in stable CAD. This lends support to the hypothesis that within the vulnerable patient exists individual vulnerable plaques with unique morphologies.

Translational Outlook

CCTA-based radiomic phenotyping of plaque in acute MI advances the role of noninvasive imaging in the detection of the rupture-prone lesions. Future studies should examine whether plaque radiomic features in patients with stable CAD can predict subsequent ACS and are modifiable by conventional or novel therapies.

Abbreviations:

ACS

acute coronary syndrome

CAD

coronary artery disease

CCTA

coronary computed tomography angiography

CP

calcified plaque

HRP

high-risk plaque

ICA

invasive coronary angiography

LDNCP

low-density noncalcified plaque

MI

myocardial infarction

ML

machine learning

NCP

noncalcified plaque