Coronary CT Angiography Radiomics for Identifying Coronary Artery Plaque Vulnerability: A Systematic Review

Abstract

Background/objectives: Acute coronary syndrome (ACS) is a major global health issue primarily caused by the rupture or erosion of vulnerable coronary plaques. Non-invasive identification of these plaques through coronary computed tomography angiography (CCTA) can be improved with radiomics, which extracts and analyses quantitative features from medical images. This systematic review aims to comprehensively evaluate the literature surrounding the role of radiomics in assessing coronary plaques via CCTA. Methods: A systematic search of Medline, EMBASE, and Web of Science was conducted up to July 2024. Nine studies met the inclusion criteria, and their methodological quality was assessed using the radiomic quality score (RQS) and the QUADAS-2 tool. Results: All studies that evaluated radiomic models for plaque vulnerability reported area under the curve (AUC) values exceeding 0.7, indicating at least modest diagnostic performance. In the four studies that made direct comparisons, radiomic models consistently outperformed conventional CCTA markers. However, RQS scores ranged from 2.7% to 41.7%, reflecting variability in study quality and underscoring the need for more robust validation. Conclusions: Radiomics has the potential to enhance CCTA-based identification of vulnerable coronary plaques, offering a promising non-invasive approach to predicting major adverse cardiovascular events. However, the current body of research is limited by the lack of external validation, reliance on small, single-centre retrospective studies, and methodological inconsistencies, which impact the generalisability and reproducibility of findings. Future research should prioritise prospective, multi-centre studies with standardised protocols and rigorous validation frameworks to effectively incorporate radiomics into clinical practice.

1. Introduction

Acute coronary syndrome (ACS), a leading cause of global morbidity and mortality, is primarily caused by rupture or erosion of vulnerable coronary artery plaques [1]. These plaques, typically large, lipid-rich atheromas with a fibrous cap, are crucial to identify for preventing ACS. Invasive imaging methods like intravascular ultrasound (IVUS) and optical coherence tomography (OCT) have pinpointed vulnerable plaque features, such as thin fibrous caps, lipid-rich necrotic cores, and microcalcifications, validated against histological standards. However, these methods are both invasive and costly [2,3,4].

Coronary computed tomography angiography (CCTA), a non-invasive imaging modality, is commonly used to diagnose coronary artery disease (CAD) and assess plaque morphology. Along with invasive imaging techniques, CCTA has identified numerous vulnerable plaque characteristics that are associated with major adverse cardiovascular events (MACE), including quantitative parameters such as low attenuation, spotty calcification, and positive remodelling and the napkin-ring sign (NRS)—a central low-attenuation area surrounded by a higher-attenuation rim [5]. However, these conventional CCTA parameters are qualitative in nature, leading to concerns of intra-observer variability and may fail to fully capture plaque complexity, heterogeneity, and latent morphological features of vulnerable plaques.

Radiomics encompasses the high-throughput extraction and analysis of a multitude of quantitative features from medical images and offers a promising approach to enhancing plaque assessment [6]. This technique enables an automated and comprehensive evaluation of plaque characteristics, potentially offering a quantitative and reproducible approach to assessing plaque vulnerability. While radiomics has demonstrated significant potential in oncology [7], its application in the field of cardiology and the assessment of coronary plaques is still emerging. Some studies suggest that CCTA-based radiomics can correlate with invasive and histologic standards of vulnerable plaques as well as identifying qualitative markers such as the NRS and may be able to predict future MACEs [4,8,9,10,11,12,13,14,15]. Given the non-invasive and widely accessible nature of CCTA, integrating radiomics to accurately and consistently identify vulnerable plaques and predict future MACEs could provide significant benefits for patients with coronary artery disease (CAD). The aim of our systematic review is to comprehensively assess the existing literature regarding the current role of radiomics in the assessment of coronary plaques on CCTA.

2. Materials and Methods

2.1. Study Design and Reporting Guidelines

This study is a systematic review of randomised and non-randomised trials and follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) reporting guidelines [16].

2.2. Search Strategy

In July 2023, we conducted searches across several databases as part of the systematic review, including Medline, EMBASE, and Web of Science. A comprehensive outline of the systematic search process, including the detailed search terms, is provided in Supplementary Material S1. The final search date was 1 July 2024. Additionally, we explored the grey literature to identify other relevant publications.

2.3. Inclusion/Exclusion Criteria

Inclusion criteria:

• Studies investigating the use of radiomics in CCTA to identify vulnerable coronary artery plaques (plaques high risk of ACS)
• Patients aged 18 years old and above
• Prospective or retrospective studies

Exclusion criteria:

• Studies not including CCTA analysis (CAC or non-contrast CT, etc.)
• Studies not investigating plaque vulnerability or solely investigating plaque progression
• Ex-vivo studies
• Case series/reports
• Consensus statements
• Conference abstracts

2.4. Study Selection, Data Extraction, and Critical Appraisal

A database was compiled using the reference management software EndNote X9™. Two researchers (HCT and CM) independently reviewed the search results. Initially, duplicates were removed. The remaining study titles were screened for potential relevance. The abstracts of studies considered potentially relevant were then reviewed for eligibility based on the predefined inclusion and exclusion criteria. Studies that did not meet the criteria were categorised in the database according to the reason for exclusion. Full texts of the eligible abstracts were further evaluated using the same criteria. Discrepancies in judgement about the potential inclusion of an article were resolved via an open discussion between the authors and an independent third reviewer (RM).

To facilitate efficient data extraction and storage, the Cochrane Collaboration’s screening and data extraction tool, Covidence, was employed [17]. Data were collected independently by two reviewers (HCT and CM) using the following categories: study details, study design, population, intervention, comparison groups, and outcomes.

A critical appraisal of the methodological quality and risk of bias of the included studies was conducted independently by two reviewers (HCT and CM). The quality assessment was performed according to the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) and Radiomics Quality Score (RQS) [18,19].

2.5. Systematic Review Registration

Our systematic review was registered on PROSPERO in August 2024 (ID: CRD42024572070).

3. Results

3.1. Search Results

The literature search, as previously described, resulted in 219 studies (see Figure 1). After removing 2 duplicates, 217 studies remained for screening. An initial screening led to the review of 99 abstracts to determine eligibility. From these, 21 studies were selected for full-text review. Ultimately, 9 studies met the inclusion criteria and were included in the analysis [4,8,9,10,11,12,13,14,15].

3.2. Methodological Characteristics and Quality of Studies

Among the included studies, eight were retrospective, while one (Lin et al. [4]) was prospective. Additionally, Chen et al. [10] incorporated a prospective analysis to validate their radiomic signature.

Table 1. Methodological characteristics of the included studies.

Author (Year)	Country	Journal	Study Design	Total Number of Patients (Number of Plaques)	Multi-Centre	Primary Outcome
Kolossváry et al. (2017) [8]	Hungary	Circ Cardiovasc Imaging	Retrospective case–control study	60 (60)	No	Identification of vulnerable plaques
Kolossváry et al. (2019) [9]	Korea	Eur Heart J Cardiovasc Imaging	Retrospective analysis	25 (44)	No	Identification of invasive and radionucleotide imaging markers of plaque vulnerability
Lin et al. (2022) [4]	Australia	JACC Cardiovascular Imaging	Post hoc analysis of prospective study	120 (120)	No	Identification of culprit lesions on CCTA
Chen et al. (2023) [10]	China	Radiology	Study 1: retrospective analysis Study 2: prospective observational study	Study 1: Development set—225 Internal test set—43 External test set—31 Study 2: 708	Yes	Study 1: Identification of vulnerable plaques Study 2: Risk of future MACE
Zheng et al. (2024) [11]	China	Coronary Artery Disease	Retrospective analysis	158	No	Identification of vulnerable plaques
Li et al. (2022) [12]	China	European Radiology	Retrospective analysis	Training group: 36 (350) Validation group: 8 (196)	No	Identification of vulnerable plaques
Yunus et al. (2022) [13]	Malaysia	Diagnostics (Basel)	Retrospective analysis	202	No	Classification of atherosclerotic plaques
Chen et al. (2021) [14]	China	European Journal of Radiology	Retrospective analysis	33 (43)	No	Identification of vulnerable plaques
Kim et al. (2023) [15]	USA	Bioengineering (Basel)	Retrospective analysis	25 (30)	No	Identification of vulnerable plaques

AMI: acute myocardial infarction, CCTA: coronary CT angiography, MACE: major adverse cardiovascular events.

provides a summary of the methodological characteristics of the included studies. The quality of the data and the risk of bias were assessed using the radiomic quality score (RQS) and QUADAS-2 tools. A detailed explanation of the tools and a breakdown of the results can be found in the Supplementary Material (S2/S3). The RQS values varied significantly among the included studies, from 2.7% to 41.7%. The QUADAS-2 results revealed that the included studies exhibited a risk of bias and raised some concerns regarding their applicability.

3.3. Acquisition Parameters

Coronary CT angiography (CCTA) was the imaging modality of choice for image acquisition and feature extraction across the included studies. The complete scanning parameters are detailed in

Table 2. Scanning parameters.

Study	Model	Additional Medications	TV	TC	ST	Matrix	Contrast
Kolossváry et al. (2017) [8]	Brilliance iCT 256, Philips Healthcare	- 0.8 mg sublingual nitroglycerin 2 min prior to image acquisition - Ivabradine/beta blocker if heart rate > 65	120 kVp	250–300 mAs	0.8 mm and 0.4 mm	512 × 512	90–100 mL of Iomeprol contrast agent
Kolossváry et al. (2019) [9]	Somatom Definition; Siemens Medical Solutions	-	-	-	-	-	-
Lin et al. (2022) [4]	Adaptive Iterative Dose Reduction 3D, Canon Medical Systems	- 0.4–0.8 mg sublingual nitroglycerin prior to image acquisition - Beta-blocker if heart rate > 65	120 kV (BMI ≥ 25 kg/m2) or 100 kVp (BMI < 25 kg/m2)	300–500 mA (depending on BMI)	0.5 mm	512 × 512	Iodinated contrast (60–90 mL, 350 mg iodine per mL [Omnipaque])
Chen et al. (2023) [10]	Somatom Definition Flash, Siemens	- 0.5 mg sublingual nitroglycerin 5 min before CCTA acquisition	70–120 kVp	-	0.75 mm	-	40–80 mL iodinated contrast media (Ultravist 370 mg iodine/mL)
Zheng et al. (2024) [11]	-	-	-	-	-	-	-
Li et al. (2022) [12]	Training group: Somatom Definition Flash, Siemens Validation group: Discovery CT 750 High Definition, GE Healthcare	-	80–120-kVp	320–400 mA (depending on BMI)	0.625 mm	-	50 to 60 mL of undiluted contrast agent (Iopromide [Ultravist] 370 mgI/mL) followed by a 30 mL 30%:70% mixture of contrast medium
Yunus et al. (2022) [13]	GE Medical System, Brand Revolution 512	-	120 kVp	Smart mA (50–430) automatic exposure control	0.63 mm	512 × 512	55–60 mL (5 mL/s) (Care bolus)
Chen et al. (2021) [14]	Somatom Definition FLASH, Siemens	- All subjects were given 0.5 mg sublingual nitroglycerin before CCTA acquisition	120 kVp	360 mAs	0.5 mm–0.75 mm	-	70 mL iodinated contrast media (iopamidol 370 mg iodine/mL, Bayer) was injected at 4–5 mL/s
Kim et al. (2023) [15]	Brilliance ICT 256 scanner, Philips Healthcare	- All subjects were given 0.5 mg sublingual nitroglycerin 5 min prior to CCTA acquisition - Target HR < 60 achieved by 100 mg metoprolol/increments of IV metoprolol 5 mg–20 mg	100–120 kVp (depending on BMI)	300–800 mAs (depending on BMI/prospective vs. retrospective gating)	-	-	80 mL iodinated contrast dye (ISOVUE 370), 5–6 mL/s

TV: Tube Voltage, TC: Tube Current, ST: Slice Thickness, CCTA: Coronary CT Angiography, BMI: Body Mass Index.

and were consistent across the studies, though these specifics were omitted in the methods of two studies [9,11].

3.4. Development of Signatures

Exact feature extraction methods varied across studies; however, a relatively similar pathway was followed across the board. Regions of interest (ROI) were segmented semi-automatically using dedicated software assisted by experienced radiologists in all but one study, which utilised manual segmentation only [12]. Radiomic features were then extracted from these segments using radiomic software packages, which varied among the included studies. The features extracted were largely grouped into three categories: first-order or intensity features, shape features, and texture features. The process of segmentation and feature processing is detailed in

Table 3. Segmentation, feature processing, and software.

Study	Segmentation Method	Segmentation Software	Validation Method	Selected Features	Radiomic Parameters Calculated	Radiomic Plaque Analysis Software	Statistical Methods
Kolossváry et al. (2017) [8]	Semi-automated	QAngioCT Research Edition, Version 1; Medis Medical Imaging Systems	5-fold cross-validation using 10,000 repeats	First-order statistics, GLCM statistics, GLRLM statistics, geometry statistics	4440	QAngioCT Research Edition Version 1 QAngioCT 3D Workbench	- NRS vs. non-NRS groups: permutation test of symmetry for matched samples using conditional Monte Carlo simulations with 10,000 replicas - Diagnostic performance estimates: AUC
Kolossváry et al. (2019) [9]	Semi-automated	QAngioCT Research Edition, Version 1; Medis Medical Imaging Systems	5-fold cross-validation with 1000 repeats	First-order statistics, GLCM statistics, GLRLM statistics, geometry statistics	935 (44 were first-order statistics; 342 were statistics calculated from GLCM; 33 were statistics extracted from GLRLM, while 516 were geometry-based)	Radiomics Image Analysis version 1	- ROC curve was calculated for each repeat and averaged for a model of diagnostic performance - AUC as a measure of diagnostic accuracy
Lin et al. (2022) [4]	Semi-automated	Autoplaque version 2.5	10 times repeated 10-fold cross-validation	First-order statistics, GLCM statistics, GLRLM statistics, geometry statistics	1103 (44 first-order metrics; 342 GLCM textural parameters; 33 GLRLM textural parameters; and 684 geometric parameters)	Autoplaque version 2.5 Radiomics Image Analysis (R.I.A version 1.4.2)	- Linear regression analysis using lesion types as categorical predictors, with stable CAD lesions as reference - Pairwise linear regression: culprit vs. nonculprit lesions - Diagnostic accuracy method: AUC
Chen et al. (2023) [10]	Semi-automated	Coronary Plaque Analysis, Syngo, version 5.0.2	10-fold cross-validated	GLCM statistics, GLRLM statistics, GLSZM statistics, GLDM, and NGTDM statistics	1691 (324 first-order features, 17 shape features, and 1350 textural features)	Radiomics, Syngo (version 1.2.1)	- Study 1: AUC, accuracy, sensitivity, and specificity of the RS and anatomic plaque parameters - Study 2: Kaplan–Meier survival curves, Anatomic plaque features, and clinical variables were both adjusted using Cox regression to determine prognostic variables
Zheng et al. (2024) [11]	Semi-automated	Darwin Intelligence Research Platform	five-fold cross-validation with 1000 replications	shape and size features, first-order statistical features, and texture features	1650 (11 shape and size figures, 204 first-order statistical features, 780 texture features)	Darwin Intelligence Research Platform. Version 1	- ROC—a model of diagnostic performance - AUC as a measure of diagnostic accuracy
Li et al. (2022) [12]	Manual	-	5-fold cross-validation	First-order statistical features, shape features, GLCM features, GLDM features, GLRLM features, GLSZM features, and NGTDM features.	1184 (234 first-order statistical features, 14 shape features, 273 GLCM features, 182 GLDM features, 208 Gy- GLRLM features, 208 GLSZM features, and 65 NGTDM features.	PyRadiomics software, version 1	- AUCs are compared using DeLong’s test. - Performance of diagnostic models, used bootstrap (1000 times) to estimate 95% Cis - Bonferroni’s correction was used for multiple-hypothesis correction
Yunus et al. (2022) [13]	Semi-automated	LIFEx (7.0.16 version)	-	first-order features, shape order features, second-order features—GLCM features, GLRLM features, NGLDM features, and GLZLM features	first-order features, 5 shape-order features, and 31 s order features	TPOT version 0.11.7	- ROC—a model of diagnostic performance - AUC as a measure of diagnostic accuracy - 4 × 4 heatmap confusion matrix
Chen et al. (2021) [14]	Semi-automated	Coronary Plaque Analysis, Syngo	Five-fold cross-validation with 100 repeats	First-order features, Second-order features, and Third-order texture features: GLDM features, GLCM features, GLRLM features, GLSZM features, and NGTDM features	1691 features extracted	Radiomics, Syngo (version 1.2.1)	- ROC—a model of diagnostic performance - AUC as a measure of diagnostic accuracy - Radiomics model was compared with conventional high-risk features by OCT with the Delong test
Kim et al. (2023) [15]	Semi-automated	Aquarius workstation	three-fold cross-validation with 1000 repetitions	Intensity-based (first order) features, shape-based features, texture features— GLDM, GLCM, GLRLM, GLSZM, and NGTDM features	1356 features were extracted (252 were shape features, 204 were intensity features, and 900 were texture features)	PyRadiomics software, version 1	- Univariate analysis to evaluate the diagnostic performance of each individual feature using ROC/AUC - Multivariate analysis to evaluate the overall diagnostic performance of the radiomics model using ROC/AUC

NRS: napkin ring sign; CAD: coronary artery disease; OCT: optical coherence tomography; ROC: receiver operating characteristics; AUC: area under curve; GLSZM: grey level size zone matrix; GLDM: grey level dependence matrix; NGTDM: neighbouring grey tone difference matrix; GLCM: grey-level co-occurrence matrix; GLRLM: grey-level run length matrix; GLSZM: grey-level size zone matrix.

.

3.5. Signature Performance

The performance of radiomic models in identifying vulnerable coronary artery plaques varied across studies, as did the results of radiomic feature assessments. Seven of the included studies developed radiomic models for assessing vulnerable plaques [4,10,11,12,13,14,15], while two studies focused solely on radiomic feature assessment [8,9]. All studies that tested a radiomic model for plaque vulnerability reported at least modest diagnostic performance (AUC > 0.7). In the four studies [4,10,12,14] that compared radiomic models to conventional CCTA markers of vulnerability, the radiomic models outperformed the traditional markers.

Table 4. Feature Assessment, Signature Performance, and Baseline Characteristics.

Study	Age (Median) [IQR] {SD}	Radiomic Feature Assessment	Radiomic Feature Assessment—Results	Outcome of Interest—Validation	Validation Results (AUC, 95% CI) [SD]
Kolossváry et al. (2017) [8]	NRS group: 63.07 (56.5–68.3) Control group: 63.9 (54.7–72.1)	Radiomic feature analysis—identifying plaques with NRS	- 440/4440 (9.9%)—AUC value > 0.80 Highest values: ○ short-run low-grey-level emphasis: 0.918 ○ long-run low-grey-level emphasis: 0.894 ○ surface ratio of component 2 to total surface: 0.890 ○ long-run emphasis: 0.888 ○ surface ratio of component 7 to total surface: 0.888	-	-
Kolossváry et al. (2019) [9]	62 (59–69)	Radiomic feature analysis—identifying attenuated plaques on IVUS	- 35/935 (3.7%)—AUC 0.70–0.79 - 311/935 (33.3%)—AUC 0.60–0.69	-	-
		Radiomic feature analysis—identifying TCFA on OCT	- 1/935 (0.1%)—AUC 0.80–0.89 - 44/935 (4.7%)—AUC 0.70–0.79 - 219/935 (23.4%)—AUC 0.60–0.69
Lin et al. (2022) [4]	Acute MI group: 59.9 {11.6} Stable CAD group: 60.2 {11.3}	Radiomic feature analysis—association with culprit lesions (linear regression analysis)	- 349/1103 (31.6%) radiomic features significantly associated with culprit lesions - 273/1103 (24.8%) radiomic features associated with all nonculprit lesions	Identification of culprit lesions (post-AMI—confirmed on angiography)	Adverse CT plaque characteristics alone: - 0.65; 95% CI: 0.57–0.74
					Quantitative plaque analysis + adverse CT characteristics: - 0.76; 95% CI: 0.68–0.83
					Radiomic features + quantitative plaque analysis + adverse CT characteristics: - 0.86; 95% CI: 0.80–0.92
					External validation cohort: - 0.84; 95% CI: 0.74–0.94
Chen et al. (2023) [10]	Study 1: Developmental set: 64 (34–89) Internal set: 65 (49–81) External set: 65 (37–81) Study 2: 62 (31–91)	Radiomic features associated with vulnerable plaques	- 861/1691 (51%)—good interobserver agreement after ICC analysis - 370/861 (43%) showed associations with vulnerable plaques	Identification of HRP (≥2 features of adverse plaque characteristics on IVUS)	Radiomic signature: Training set: - 0.81; 95% CI: 0.75, 0.87 Validation set: - 0.75; 95% CI: 0.64, 0.85 Internal test set: - 0.80; 95% CI: 0.64, 0.97 External test set: - 0.77; 95% CI: 0.60, 0.94
					HRP alone: - AUC, 0.58; 95% CI: 0.54, 0.62
					HRP + quantitative plaque parameters: - 0.73; 95% CI: 0.67, 0.80
					RS + HRP + quantitative plaque parameters: - 0.81; 95% CI: 0.75, 0.87
Zheng et al. (2024) [11]	Non-high risk: 62.6 {11.35} High risk: 66.4 {9.04}	-	-	Performance in predicting vulnerable plaques (coronary artery plaque causing AMI within 2 years of follow-up)	Training cohort: - 0.977
					Test cohort: - 0.815
Li et al. (2022) [12]	Training group: 53.0 {9.0} Validation group: 48.5 {11.6}	-	-	Training group diagnostic accuracy—identifying vulnerable plaques (based on histological characteristics)	- NB: 0.899 [±0.035]
					- DT: 0.846 [±0.046]
					- RF: 0.892 [±0.033]
					- GBDT: 0.869 [±0.044]
					- SVM: 0.899 [±0.031]
					- MLP: 0.871 [±0.039]
					- LR: 0.90 [±0.033]
					- K-NN: 0.866 [±0.032]
				Validation group diagnostic accuracy—identifying vulnerable plaques (based on histological characteristics)	- SVM: 0.754, 95% CI: 0.675–0.830
					- LR: 0.770, 95% CI: 0.702–0.846
					- NB: 0.782, 95% CI: 0.710–0.846
					- Conventional CCTA model: 0.656, 95% CI: 0.593–0.718
Yunus et al. (2022) [13]	58.84 {9.49}	-	-	Classification of plaques—calcified, mixed plaques, normal arteries	Model 1: - Normal: 0.91 - Calcified: 0.83 - Non-calcified: 0.69 - Mixed: 0.77
					Model 2: - Normal: 0.86 - Calcified: 0.78 - Non-calcified: 0.72 - Mixed: 0.74
					Model 3: - Normal: 0.50 - Calcified: 0.68 - Non-calcified: 0.58 - Mixed: 0.62
					Model 4: - Normal: 0.93 - Calcified: 0.82 - Non-calcified: 0.69 - Mixed: 0.76
Chen et al. (2021) [14]	59 (41–88)	Association of radiomic features with TCFA seen on OCT	- 35/1691 radiomic features—showed a significant difference between plaques with or without TCFAs (All p < 0.05)	Identification of TCFA lesions (OCT as standard)	Radiomics model: - 0.952; 95% CI 0.897–1.000
					HRP features model: - 0.621; 95% CI: 0.469–0.773
					FAI model: - 0.52; 95% CI: 0.33–0.70
Kim et al. (2023) [15]	63.0 {11}	Association of radiomic features with vulnerable plaque characteristics on IVOCT:	293 PCAT-LOI-radiomics features—identification of IVOCT-TCFA: - AUC 0.70–0.79: 2% (6/293) - AUC 0.60–0.69: 17.1% (50/293) - AUC 0.50–0.59: 31.1% (91/293)	Identification of TCFA lesions (OCT as standard)	PCAT-Vessel radiomics model: - 0.77 [±0.14]
					PCAT-LOI radiomics model: - 0.78 [±0.13]
			341 PCAT-Vessel-radiomics features—identification of IVOCT-TCFA: - AUC 0.70–0.79: 3% (3/341) - AUC 0.60–0.69: 26.3% (26/341) - AUC 0.50–0.59: 70.7% (238/341)	Identification of MC lesions	PCAT-vessel radiomics model: - 0.89 [±0.09]
					PCAT-LOI radiomics model: - 0.83 [±0.12]
			293 PCAT-LOI-radiomics features—identification of IVOCT-TCFA: - AUC 0.70–0.79: 1% (3/293) - AUC 0.60–0.69: 21.8% (64/293) - AUC 0.50–0.59: 18.1% (53/293)	Identification of combined TCFA and MC lesions	PCAT-vessel radiomics model: - 0.91 [±0.09]
			341 PCAT-Vessel-radiomics features—identification of IVOCT-TCFA: - AUC 0.70–0.79: aomos 1.5% (5/341) - AUC 0.60–0.69: 10.9% (37/341) - AUC 0.50–0.59: 23.2% (79/341)		PCAT-LOI radiomics model: - 0.88 [±0.10]

AUC: area under curve; CI: confidence interval; NRS: napkin-ring sign; IVUS: intra-vascular ultrasound; ICC: intraclass correlation coefficient; AMI: acute myocardial infarction; NB: naïve bayes; DT: decision tree; RF: random forest; GBDT: gradient boosting decision tree; SVM: support vector machine; MLP: multi-layer perceptron; LR: logistic regression; K-NN: K-nearest neighbours; HRP: high risk plaque; FAI: fat attenuation index; OCT: optical coherence tomography; IVOCT: intravascular optical coherence tomography; LOI: Lesion region Of Interest; PCAT: pericoronary adipose tissue; MC: microchannels; TCFA: thin-cap fibroatheroma.

details the performance of each model in predicting primary outcomes as well as the results of radiomic feature assessments. Chen et al. [10] were the only study to also include a prospective analysis to assess the prognostic value of their radiomic model. They found that a high radiomic signature was independently linked to MACE during a median follow-up of 3 years (hazard ratio: 2.01, p = 0.005).

4. Discussion

Our review highlights the current state of the literature on the development and performance of radiomic models for assessing vulnerable coronary artery plaques. The optimal signatures developed within the included studies predicted their primary outcome with at least modest accuracy (AUC range 0.72–1) and consistently outperformed conventional CT markers of plaque vulnerability in the four studies [4,10,12,14] where this comparison was made. Additionally, all the studies demonstrated that radiomic analysis significantly enhances the information that can be extracted from CCTA. This highlights the potential of incorporating radiomics into conventional CT coronary angiography, offering a non-invasive and widely accessible imaging technique capable of providing a quantitative and reproducible assessment of plaque vulnerability. By extracting detailed imaging features, this approach can offer valuable insights into plaque morphology, potentially enhancing the prediction and prevention of MACE, as demonstrated in the prospective trial by Chen et al. [10].

The RQS was employed to assess the methodological quality of the radiomics-based investigations used in these studies, to identify high-quality results as well as issues limiting their value and applicability. The RQS scores ranged widely from 2.7% (1/36) to 44.4% (16/36), with a mean value of 8.77 (24.4%) (Supplementary Material S1), which aligns closely with the findings of a systematic review by Lee et al., where the quality of radiomic research in cardiac CT was reported to have a similar value of RQS (9.9, 27.4%) [20]. This similarity suggests that these methodological issues are prevalent across the field. Many studies scored poorly due to a lack of multivariable analysis, retrospective design, and particularly, the absence of multi-centre validation, which limits the generalisability of the findings. Prospective validation, preferably conducted in a multi-centre setting, is essential for the future clinical application of radiomics, as it addresses variability and confirms the model’s predictive accuracy in real-world clinical practice [21]. Additionally, the QUADAS-2 score raised concerns about potential bias in patient selection and the applicability of the index test. Future studies should apply rigorous methodological standards and adopt more comprehensive risk-of-bias assessments to minimise bias and ensure the reliability of their results.

Our review is limited by the relatively small number of studies and their modest sample sizes, a challenge that is unavoidable given the emerging nature of this field. The heterogeneity in radiomic workflows, study methodologies, and primary outcomes prevented us from conducting a meta-analysis of the included studies. To provide a comprehensive overview of the current research landscape, we included all studies that met the inclusion criteria, irrespective of their quality. This approach allows us to highlight the strengths and weaknesses of existing research and to offer recommendations for future studies.

The primary limitations identified within the reviewed studies include the small sample sizes, single-centre designs, and retrospective methodologies, which contribute to potential overfitting and selection bias and limit the generalisability of the findings. To address the issue of small sample sizes, the researchers employed stratified cross-validation with numerous repeats to generate reliable estimates of diagnostic performance. While this approach is widely recognised as a method for assessing the generalisability of diagnostic accuracy in real-world scenarios, it cannot fully replace the external validation of results in significantly larger and independent cohorts [22]. The generalisability of these results is also influenced by differences in CT scanners from various vendors, as well as inconsistencies in image acquisition, feature extraction, and data analysis techniques. Multiple studies have shown that CT scanner variability, reconstruction settings, and segmentation methods can significantly affect radiomic signatures, underscoring the need for standardised protocols and ensuring a robust framework for radiomic analysis that can facilitate cross-study comparisons [23,24]. For instance, Chen et al. [10] showed that only half of the 1691 initially extracted radiomics features exhibited strong interobserver agreement after intraclass correlation coefficient analysis. To address these challenges, the European Society of Radiology (ESR) has introduced the European Imaging Biomarkers Alliance (EIBALL) with the goal of improving the standardisation and accuracy of quantitative imaging parameters, thereby ensuring consistent and reproducible results across different centres and imaging platforms.

Another limitation among the included studies was the inconsistent criteria used to define a vulnerable coronary artery plaque. The definitions varied across studies, with some relying on qualitative CCTA markers of vulnerability like the NRS, while others used optical coherence tomography (OCT), intravascular ultrasound (IVUS) markers, histological characteristics, or plaques associated with acute myocardial infarction (AMI). Although these markers are linked to plaque vulnerability and the risk of MACE, the variability in primary outcomes diminishes the overall generalisability of the findings, particularly when qualitative markers with intra-observer variability are employed [25,26].

The included studies highlighted several other challenges that limit the integration of CCTA radiomics into routine clinical practice. These include technical complexity, the absence of user-friendly automated software, and the labour-intensive nature of the current semi-automated segmentation process, which also carries the risk of inter-observer variability. Unlike oncological radiomics, where segmentation along tumour borders is straightforward, cardiovascular radiomics faces difficulties in accurately segmenting small, geometrically complex regions of cardiac plaque [27]. All studies noted the significant workload associated with coronary plaque radiomic analysis, which currently hinders its practical use. The eventual aim is to develop fully automated, user-friendly, artificial intelligence (AI)-based radiomic plaque analysis, which can be embedded into routine CCTA reporting software, providing real-time evidence regarding the vulnerability of coronary plaques and the risk of future AMI. Advances in imaging technologies, such as photon-counting detector CT, may facilitate automated segmentation by improving signal-to-noise ratio and spatial resolution, reducing the workload and variability associated with manual processes [28].

In recent years, CCTA has been widely used due to its non-invasive nature, cost-effectiveness, and accessibility, coupled with its excellent negative predictive value for ruling out obstructive coronary artery disease [29]. The primary disadvantage of CCTA is exposure to ionising radiation. To address this concern, innovative non-invasive diagnostic modalities, such as the modified Haller index (MHI), have been introduced to reduce the number of unnecessary procedures performed in patients with a low probability of CAD [30]. Integrating radiomics with CCTA does, however, hold significant promise for enhancing current imaging workstations, enabling the rapid and accurate identification of vulnerable plaques without the need for invasive angiography, and helping to predict and prevent major adverse cardiovascular events (MACE). Additionally, radiomic analysis of coronary plaques may offer insights into the underlying biology of these vulnerable, and monitoring the evolution of these radiomic features over time could help evaluate the effectiveness of therapies targeting coronary artery plaques.

To effectively translate radiomics into the field of cardiac imaging, we must first address the challenges of variability in feature extraction and the lack of reproducibility, which are significant barriers to the success of future radiomic-based research. Additionally, large-scale, prospective, multi-centre studies with diverse patient populations, standardised protocols, and robust external validation frameworks are essential to improving the generalisability of prediction models. Such advancements require collaboration among academic institutions. The Oxford Risk Factors and Non-Invasive Imaging (ORFAN) study currently underway in the UK is one example of a trial working toward these goals.

5. Conclusions

Our review highlights the growing potential of radiomic models in assessing vulnerable coronary artery plaques, demonstrating that these models can predict plaque vulnerability with moderate to high accuracy and surpass conventional CT markers. Radiomic analysis offers a promising non-invasive approach for detailed plaque evaluation and MACE prediction. However, the current literature is limited by a lack of external validation, small single-centre retrospective studies, and inconsistent methodologies that lead to variations in feature extraction, ultimately affecting the generalisability and reproducibility of the findings. Addressing these issues will require large-scale, multi-centre studies with standardised protocols and robust validation frameworks to successfully integrate this technology into clinical practice.