Development of a clinical prediction model for acute myocarditis using coronary computed tomography angiography-based radiomics

Introduction

Acute myocarditis is a common cardiovascular disease in young individuals, with an incidence rate of approximately 4–14 per 100,000 people annually and a mortality rate of 1–7% (1). Early detection and intervention can improve the prognosis of patients with acute myocarditis. Coronary computed tomography angiography (CCTA) is generally not used for diagnosing acute myocarditis. Early studies have demonstrated that CCTA can reveal some cases of myocarditis (2,3). The latest spectral cardiac computed tomography (CT) can detect an elevated extracellular volume (ECV) in patients with acute myocarditis (4,5). However, the severity of myocardial injury can vary, patients with a mild injury may be less likely to be visualized on CCTA.

Radiomics is the conversion of images into higher dimensional data, and the subsequent data mining to better support the clinical decision (6). CCTA is mostly used for the detection of coronary artery disease, and radiomics derived from CCTA have mostly focused on vascular, plaque, and pericardial fat in recent years (7-10). There are only a few CCTA-based radiomic studies on myocardium. Two studies (11,12) demonstrated that CCTA could help differentiate between normal and infarcted myocardium and between acute and chronic myocardial infarction. In one study with a small sample size (13), texture analysis of CCTA-based radiomics for acute myocarditis demonstrated that the texture of necrotic myocardium differed from that of normal myocardium. Currently, there is a lack of CCTA-based radiomics studies with a large sample size for acute myocarditis, including studies that have evaluated the difference of CCTA-based radiomics in patients with acute myocarditis and normal individuals, because both acute myocarditis patients and normal cohort usually present with normal CCTA performance, and the performance of CCTA radiomics on the prediction for myocarditis is still unclear. Moreover, further assessing the use of radiomics in distinguishing fulminant from non-fulminant myocarditis, and its performance in prognosis evaluation is unknown.

In our study, we aimed to use CCTA-based radiomics with machine learning to construct a clinical prediction model for acute myocarditis and explore the distinguishing performance of radiomics in patients with or without fulminant myocarditis and major adverse cardiovascular events (MACEs). We present this article in accordance with the TRIPOD reporting checklist (available at https://cdt.amegroups.com/article/view/10.21037/cdt-24-330/rc).

Methods

Patients and study design

In this study, we retrospectively included patients with myocarditis who presented to the Affiliated Jinhua Hospital, Zhejiang University School of Medicine (Center 1) and Sir Run Run Shaw Hospital, Zhejiang University School of Medicine (Center 2) between January 2013 to October 2023. Patients diagnosed with acute myocarditis following a CCTA examination, which was confirmed by cardiac magnetic resonance (CMR), were included in the study. All study patients met the 2009 or 2018 CMR-based Lake Louise Criteria (LLC) for diagnosing myocarditis (14,15). The exclusion criteria were as follows: CCTA slice thickness of ≥1 mm; an interval of >2 weeks between CCTA and CMR; poor quality of CCTA or CMR; presence of other cardiac diseases such as coronary artery disease with coronary stenosis of >50%, myocardial infarction, cardiomyopathy, congenital heart disease, valvular disease, and metabolic diseases involving the heart. Patients who presented with rapid and severe hemodynamic disturbances, including acute heart failure, hypotension, shock, electric instability, or rapidly evolving conduction abnormalities, were considered to have fulminant myocarditis (16,17). Patients with a normal CCTA and CMR were included in the control group. A total of 86 patients with acute myocarditis and 129 patients with normal CCTA were included in the study (Figure 1). Follow-up data was obtained via examination of electronic medical records, direct patient communication, or via telephone interview with the patient or their family members. The following data were obtained from electronic medical electronic records or via direct or telephone interviews: CCTA and CMR imaging data, baseline characteristics. MACE included all-cause mortality, ventricular fibrillation or tachycardia, implantable cardioverter defibrillator discharge, and hospitalization due to progression of heart failure. Figure 2 shows the workflow of model establishment in this study. This study was conducted in accordance with the principles of the Declaration of Helsinki (as revised in 2013) and was approved by the Affiliated Jinhua Hospital, Zhejiang University School of Medicine Ethics Committee (No. 2024-70) and Sir Run Run Shaw Hospital, Zhejiang University School of Medicine Ethics Committee (No. 2024-280). Written informed consent was obtained from each patient.

Figure 1 Flow chart of the recruitment and exclusion of study patients. CCTA, coronary computed tomography angiography; CMR, cardiac magnetic resonance; LV, left ventricular; CHD, coronary heart disease; MI, myocardial infarction; ECG, electrocardiogram; ECHO, echocardiography.

Figure 2 Workflow of model establishment. 3D left ventricular myocardium segmentation was automatically performed using CoronaryDoc^®. LASSO and 5-fold cross-validation were used to evaluate results and select the optimal regularization parameter alpha. The best radiomics prediction model was selected and combined with clinical labels to construct a clinical-radiomics model for classifying patients as with or without myocarditis. 3D, three-dimensional; LV, left ventricular; LASSO, least absolute shrinkage and selection operator.

CT acquisition

Before obtaining a CCTA, the heart rate was checked and breathing exercises were performed at both study centers. Patients with a heart rate of >80 beats/min were administered 25 mg of a β-blocker (metoprolol tartrate, AstraZeneca, Wuxi, China). If the heart rate remained unsatisfactory after half an hour, the CCTA was rescheduled. The patients were trained to hold their breath for 10 seconds at the end of inspiration.

At Center 1, CCTA was performed using a 256-slice spiral CT (iCT, Philips, Best, Netherlands). A retrospective electrocardiogram (ECG)-gated scan was obtained in patients with a heart rate of >60 beats/min, and prospective ECG-triggered scan was obtained in patients with a heart rate of <60 beats/min. The tube voltage was set at 120–140 kV, the current at 900–1,200 mAs, and reconstructed slice thickness at 0.625 mm. The contrast agent, ioversol (350 mg/100 mL, Hengrui Pharmaceuticals, Lianyungang, China), was injected via the median cubital vein at a dosage of 0.8–1.0 mL/kg. Subsequently, 40 mL of saline was injected at a rate of 5.0 mL/s. The CCTA was performed using bolus-tracking technique, with the descending aorta set as the region of interest, a CT value of 120 Hounsfield units (HU), and a delay of 5.9 s after the contrast injection. The area from the tracheal bulge to the subdiaphragm was scanned.

At Center 2, CCTA was performed using a dual-source 64-detecter spiral CT (SOMATOM Definition Flash, Siemens, Erlangen, Germany). A retrospective ECG-gated scan was performed in patients with a heart rate of >60 beats/min, and a high-pitched prospective ECG-triggered scan (flash scan) was performed in patients with a heart rate of <60 beats/min. The tube voltage was set as 120 kV, the current at 150–400 mAs, and the reconstructed slice thickness at 0.75 mm. The contrast agent used was iopromide (Ultravist, 350 mg/100 mL, Bayer Scherig Pharma AG, Guangzhou, China), and the contrast dosage, injection rate, saline dosage, and scan coverage were consistent with those of Center 1. The bolus-tracking technique was used, with the aortic root set as the region of interest, a CT value of 100 HU, and a post-contrast injection delay of 5 s. For the exposition dose, the dose of Philips, Siemens, and GE scanner was between 900–1,700, 300–700, 200–600 dose-length product (DLP) (mGy·cm), respectively.

Left ventricular myocardium segmentation

Three-dimensional (3D) left ventricular myocardium segmentation of the CCTA images was performed using CoronaryDoc^® (version 6.18; Shukun Technology Co., Ltd., Beijing, China) and a 3D U-Net-based deep learning model. We adapted our deep learning model with additional attention mechanism, such as (squeeze and excitation) SE blocks, to automatically focus on myocardial regions. The application of multiscale dilated convolutions provided our model with receptive fields of different sizes, which further improved our deep learning to accurately segment different-sized anatomical structures. To ensure the accuracy of the segmented regions of the left ventricular myocardium, two senior radiologists reviewed and adjusted the regions. The segmented left ventricular myocardium was used as the region of interest in the follow-up study.

Feature selection and radiomics signature construction

The CT and region of interest images were uploaded to CoronaryDoc^®. Features were selected based on the training cohort. The extracted features were initially scrutinized for any missing values, which were subsequently imputed with the mode. Following imputation, the features were standardized using z-score normalization in order to eliminate differences caused by various CT machines and ensure the consistency during the training process. Pearson correlation analysis was used to eliminate redundant features with a threshold value of 0.9. Least absolute shrinkage and selection operator (LASSO) regression with 5-fold cross-validation was employed to identify the optimal regularization parameter, effectively balancing model complexity with its capacity to generalize to new datasets. This rigorous validation process facilitated the selection of prognostic features and clinical features, which are essential for constructing a robust radiomics signature. A Rad-score was computed for each patient using a linear combination of selected features weighted by their respective coefficients. The selected features were also used to test the performance of the model in patients with fulminant myocarditis and those with MACE.

Model building and clinical application

The dataset was randomly divided into training and test cohorts at a ratio of 7:3. In the training cohort, the patients were classified into two groups according to the Rad-score (cut-off value =0.609): with myocarditis and without myocarditis. The top 10 radiomics features that were extracted were used to construct models using the following classifiers: logistic regression, Adaboost, support vector machine (SVM), XGBoost, passive aggressive, perceptron, K-nearest neighbors, linear support vector classification (linearSVC), random forest, and decision tree. Five-fold cross-validation was used to avoid overfitting. The best radiomics prediction model was chosen and combined with the clinical labels to construct the clinical-radiomics model for classifying patients with and without myocarditis.

Statistical analysis

All data were analyzed using SPSS (version 28.0; SPSS Inc., Chicago, IL, USA) and R (version 4.3.0; https://www.r-project.org/). The continuous variables were presented as mean ± standard deviation. The categorical data were compared using the Chi-squared or Fisher’s exact test. The Kolmogorov-Smirnov test was used to assess the data distribution. The independent samples Student t-test and the Mann-Whitney U test were used to compare the normally distributed and non-normally distributed data, respectively between groups. The discriminatory power of the radiomics features for fulminant and non-fulminant myocarditis was evaluated using area under the curve (AUC) of receiver operator characteristic (ROC) curve. Univariate Cox proportional hazards regression analysis was used to calculate the hazard ratio and 95% confidence interval. The Delong test was used to compare the diagnostic efficiencies of the different radiomic models. Sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), accuracy, and F1 score were calculated using the confusion matrix. Calibration curves were drawn and decision curve analyses were performed using R. The Hosmer-Lemeshow test was used to evaluate the statistical differences between the predicted and actual probabilities. Differences were considered statistically significant at P<0.05.

Results

Patient characteristics

Seventy-three patients with acute myocarditis and 109 patients with normal CCTA from Center 1, and 13 patients with acute myocarditis and 20 patients with normal CCTA from Center 2 were included in the study. The baseline information is shown in Table 1 and 7 clinical parameters were included in the model, including age, chest pain, baseline disease, brain natriuretic peptide (BNP), echocardiography, prodromal respiratory infection, and cardiac troponin. Eight patients met the criteria for fulminant myocarditis. MACE was recorded in nine patients, and it included all-cause mortality (n=1) and hospitalization due to progression of heart failure (n=8). In this study, 86 patients with acute myocarditis and a CCTA examination were included. None of the patients demonstrated abnormal myocardial density on CCTA, and None of the patients demonstrated abnormal myocardial density or features suggestive of myocarditis on CCTA. However, a retrospective review of the CCTA revealed that 21 patients (24.42%) demonstrated a small pericardial effusion. A retrospective comparison of the CMR and CCTA revealed that 25 patients (29.07%) exhibited a mild decrease in the left ventricular myocardium density during the CCTA’s arterial phase.

Model establishment

A total of 1,173 radiomics features were extracted for each patient, including shape, first-order, second-order features, and texture features. A Pearson correlation matrix was established, and paired feature correlation coefficients were calculated. Feature pairs with an absolute Pearson correlation coefficient of >0.9 were eliminated. Of the 1,173 features, 349 were selected for subsequent analysis (Figure 3). LASSO and 5-fold cross-validation were used to validate the results and select the best regularization parameter alpha (Figure 4). A Rad-score was computed for each patient using a linear combination of selected features weighted by their respective coefficients. LASSO analysis revealed 10 radiomics features and seven significant clinical features which were included in the final analysis (Figure 5). Gray level co-occurrence matrix (GLCM) features include wavelet-LHL glcm DifferenceEntropy, wavelet-LLH glcm JointEnergy, and wavelet-HHL glcm Autocorrelation. First-order features include 1bp-3d-m2 firstorder RootMeanSauared, exponential firstorder Maximum, log-sigma-3-0-mm-3D firstorder Kurtosis, logarithm firstorder Kurtosis, and wavelet-HHL firstorder Entropy. Gray level run length matrix (GLRLM) features include wavelet-LLH glrlm RunVariance and wavelet-HHL glrlm HighGravlevelRunEmphasis.

Figure 3 A Pearson correlation matrix of radiomic features. The color bar refers to the Pearson coefficient between radiomics features.

Figure 4 Coefficients of radiomics features and clinical parameters in the LASSO model. BNP, brain natriuretic peptide; ECHO, echocardiography; cTnI, cardiac troponin I; LASSO, least absolute shrinkage and selection operator.

Figure 5 Rad-scores of training and test datasets.

Model performance

The ROC curves of the three models using SVM performed the best using training and test datasets (Table 2, Figure 6). Models 1, 2, and 3 were Rad-score model, clinical factors model, and clinical factors and Rad-score model. Table 3 shows the model performance parameters, such as AUC, accuracy, sensitivity, specificity, PPV, and negative predictive value (NPV), in detail. Model 3 achieved the most encouraging results, with AUCs of 1.000 (0.999–1.000) and 0.951 (0.880–1.000) in the training and test datasets, respectively. Using Model 1, the AUCs were 0.970 (0.949–0.991) and 0.912 (0.832–0.992) in the training and test datasets, respectively. In Model 2, the AUCs were 0.992 (0.983–1.000) and 0.943 (0.875–1.000) in the training and test datasets, respectively. A confusion matrix (Figure 7) was obtained for the training and test datasets on the basis of the classification provided by the radiologists and the models. Model 3 yielded an excellent performance for the prediction of acute myocarditis in patients with a normal CCTA.

Figure 6 ROC curves of using the training and test datasets in (A) Model 1 (Rad-score model), (B) Model 2 (clinical factors model), and (C) Model 3 (clinical factors and Rad-score model). SVM, support vector machine; AUC, area under the curve; CI, confidence interval; ROC, receiver operator characteristic.

Figure 7 Confusion matrix used for the training and test datasets based on the classification provided by the radiologists and Model 3 (clinical factors and Rad-score model). SVM, support vector machine.

Radiomics performance in identifying fulminant myocarditis and patient prognosis

The performance of radiomics features in patients with fulminant myocarditis and MACE is illustrated in Table 4. In patients with fulminant and non-fulminant myocarditis, auto__log-sigma-3-0-mm-3D_firstorder_Kurtosis, and auto__wavelet-LHL_glcm_DifferenceEntropy reached statistical significance. The ROC analysis revealed poor performance of radiomics features. In patients with and without MACE, auto__log-sigma-3-0-mm-3D_firstorder_Kurtosis, auto__wavelet-HHL_glcm_Autocorrelation, and auto__logarithm_firstorder_Kurtosis reached statistical significance. No parameter demonstrated statistical significance in the univariate Cox proportional hazards regression analysis.

Discussion

In this study, the combined clinic model based on CCTA radiomics and clinical data was proved to be with better performance than Rad-score model and clinical factors Model, and the clinical prediction performance was excellent. We prove the clinical feasibility in diagnosing myocarditis and its potential clinical value.

In previous reports (2,3) of connective tissue disease-associated myocarditis and acute myocarditis, the arterial phase of CCTA demonstrated an area of low attenuation suggestive of edema. However, the sample size in these two studies was limited. In our study, only 25 of the 86 cases demonstrated decreased myocardial density in the arterial phase. These patients who were previously considered normal demonstrated an abnormal CCTA now. Retrospective analysis of CMR and CCTA regions corresponding to edematous and necrotic segments on CMR revealed that the decreased density of the injured myocardium on CCTA when compared with the normal myocardium. The changes were visualized after adjusting the window width and for the CT values of different myocardial segments. In these patients, the decrease in density was mild in the CT range of 20 to 50 HU. When combined with delayed CT, conventional CCTA will be provided more diagnostic data (18). Our study used conventional CCTA without delayed enhancement scans. Thus, a comparison of the characteristics of late iodine enhancement (LIE) could not be assessed. ECV is demonstrated elevated in patients with myocarditis (19) and could be used to distinguish acute myocarditis from a normal myocardium (5). Dual-energy CT (DECT) is not yet routinely used in clinical practice for cases of acute myocarditis. This prolongs the scanning time, increases the radiation dose, and has a post-process that may be complex for some CTs. Thus, the use of CT-ECV is limited.

CCTA-based radiomics has gained importance in recent years. The progress on coronary calcification, plaque, pericoronary fat, and flow reserve fraction can assist in risk assessment of acute coronary syndromes, prediction of vulnerable plaques and MACEs, differentiation of myocardial infarction from angina pectoris, and identification of myocardial ischemia, scar, prognosis and non-ischemic diseases (7-10,20-23). Analysis of myocardial texture on CT has revealed that kurtosis and short run high gray emphasis demonstrated high efficacy in discriminating acute infarcted myocardium from normal myocardium (11), normal myocardium from acute and chronic myocardial infarction (12), and increased the chances of identifying myocardial ischemia when compared with the traditional model (24).

The use of machine and deep learning on clinical database has demonstrated a significant diagnostic and predictive value for acute myocarditis (25,26). Radiomic studies on echocardiography have demonstrated an increase in the texture parameters entropy, angular second moment, and long-term emphasis (27), and a decrease in the inverse difference moment and run-length nonuniformity (28). Radiomic studies on CMR cine and T2-weighted images demonstrated differences in texture characteristics between edema and normal myocardium (29,30), myocardial infarction from myocarditis (31). When using T2-mapping texture features to build radiomics model, the diagnostic performance was higher than the LLC (32). Furthermore, the combined model of mapping parameters and radiomics features has demonstrated a good predictive value for heart failure in patients with myocarditis (33).

There were few studies using radiomics on CCTA of patients with acute myocarditis. In one study (13), seven patients with post-myocarditis scars and implantable cardioverter defibrillator were enrolled. This study’s results demonstrated that the energy, entropy, kurtosis, mean, and root mean square error that were calculated using the angiographic images were significantly different between normal and scarred tissue. Other important parameters were the mean, median, and root mean square error using the ECV map. The model using the energy parameter in the baseline scan demonstrated the highest accuracy (0.94±0.17). Although the study initially demonstrated the feasibility of CCTA texture analysis for acute myocarditis, its performance was limited by the small sample size and there was lack of a biopsy or CMR for diagnostic confirmation. All patients underwent a defibrillator implantation after acute myocarditis, which suggests that these patients may have demonstrated a low myocardial density, LIE, and increased ECV. In our study, the sample size was larger and included patients with fulminant, moderate, and mild myocarditis. In our study, all patients met the 2009 or 2018 CMR-based LLC for myocarditis, which is more reliable than dual-energy CCTA findings. Hence, our results may be more representative of the condition than CCTA findings. To the best of our knowledge, this is the first study on machine learning using CCTA-based radiomics for acute myocarditis, which included data from two research centers. The cohort included 86 patients with myocarditis and 129 controls, and the predictive model included the energy, radiomics and clinical features. Machine learning was used for CCTA data mining to construct a prediction model, which demonstrated a good predictive efficacy and reflected the additional value of a conventional CCTA in the diagnosis of myocarditis. Because of the limited myocarditis-related data in this study, a bigger sample size in future studies is required to explore the value of CCTA-based radiomics in predicting the prognosis of myocarditis.

In our prediction model, five patients were misclassified. This may be attributed to the fact that the CCTA of most patients was reconstructed using the diastolic phase images. However, images of only few patients used systolic phase images. Furthermore, the myocardium segmentation was not exactly the same between systolic and diastolic phases. The inconsistency in image quality may have also contributed to the misclassification. This may be attributable to the inconsistent contrast enhancement in different individuals, especially in patients with heart failure, the insufficient enhancement of cardiac chambers, and residual contrast agent in the right atrium and right ventricle. The small sample size may have affected the prediction accuracy of the model and resulted in the misclassification of patients. Finally, the different scanner and protocols may have attributed to the misclassification, which may have resulted in a bias.

The auto__log-sigma-3-0-mm-3D_firstorder_Kurtosis, auto__wavelet-LHL_glcm_DifferenceEntropy, auto__wavelet-HHL_glcm_Autocorrelation, and auto__logarithm_firstorder_Kurtosis were valuable in patients with fulminant myocarditis and MACE. However, the discriminatory power in patients with and without fulminant myocarditis was poor. Furthermore, the performance of radiomics features in determining prognosis, as assessed by univariate Cox proportional hazards regression analysis, was also poor. These findings may be attributable to the small sample size of patients with fulminant myocarditis and MACE.

Limitations

There are some limitations in this study. First, the data were collected from two hospitals, which were mixed for constructing the model. Therefore, external validation studies are needed to improve the generalizability of the model. Originally, data from Center 1 was planned for the internal training and testing and data from Center 2 was planned for external validation. However, statistical analysis revealed that the diagnostic performance of the training and internal testing set was excellent, but that of the external validation set was low. After Pearson’s correlation and LASSO regression analysis, 11 of the 1,169 radiomic features (7 first-order and 4 second-order features), including two GLDM and two GLRLM features were selected to construct our prediction model, which demonstrated a high capacity of discrimination (AUC training and test sets, 0.998 vs. 0.908) in 184 consecutive patients (training, n=128; test, n=56). The AUC using the external test dataset was 0.782. This may be attributable to the small sample size and the different scanners and protocol used in Center 2. We believed that mixing the data from two centers can minimize the bias from data obtained from different scanners, and this may have negated the bias produced by the small sample size of the external dataset. Second, the difference in the reconstructed myocardium between patients, majority of which were reconstructed during the diastolic phase, was inconsistent. The texture characteristics in CMR in different cardiac phases are different (34). However, it is unclear whether these differences are visualized on CCTA. Third, there were still some potential biases, due to small sample size from two centers, inconsistent data sources in which most data were acquired from Philips scanner and a few were acquired from Siemens and GE scanners. Furthermore, acute myocarditis was a clinical diagnosis, which was not confirmed by endomyocardial biopsy. Although CCTA radiomics is of limited clinical applicability, we believe that multi-center and prospective data validation is important and promising in future work.

Conclusions

The combined model based on CCTA-based radiomics and clinical data demonstrated excellent performance. The radiomics features demonstrated a difference between patients with and without fulminant myocarditis, and between patients with and without MACE. However, its prognostic significance remains limited.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://cdt.amegroups.com/article/view/10.21037/cdt-24-330/rc

Data Sharing Statement: Available at https://cdt.amegroups.com/article/view/10.21037/cdt-24-330/dss

Peer Review File: Available at https://cdt.amegroups.com/article/view/10.21037/cdt-24-330/prf

Funding: This study was supported by the Medical and Health Research Project of Zhejiang Province (No. 2025KY1745) and the project grant from the Jinhua Municipal Central Hospital (No. JY2023-2-02).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://cdt.amegroups.com/article/view/10.21037/cdt-24-330/coif). L.L., D.G., and Y.H. are current employees of ShuKun Technology Co., Ltd. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the principles of the Declaration of Helsinki (as revised in 2013) and was approved by the Affiliated Jinhua Hospital, Zhejiang University School of Medicine Ethics Committee (No. 2024-70) and Sir Run Run Shaw Hospital, Zhejiang University School of Medicine Ethics Committee (No. 2024-280). Written informed consent was obtained from each patient.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ammirati E, Moslehi JJ. Diagnosis and Treatment of Acute Myocarditis: A Review. JAMA 2023;329:1098-113. [Crossref] [PubMed]
2. Brett NJ, Strugnell WE, Slaughter RE. Acute myocarditis demonstrated on CT coronary angiography with MRI correlation. Circ Cardiovasc Imaging 2011;4:e5-6. [Crossref] [PubMed]
3. Tan S, Moir S, Seneviratne S. Connective Tissue Disease Associated Myocarditis on Computed Tomography Coronary Angiogram. Heart Lung Circ 2021;30:e121-2. [Crossref] [PubMed]
4. Si-Mohamed SA, Restier LM, Branchu A, et al. Diagnostic Performance of Extracellular Volume Quantified by Dual-Layer Dual-Energy CT for Detection of Acute Myocarditis. J Clin Med 2021;10:3286. [Crossref] [PubMed]
5. Si-Mohamed SA, Congi A, Ziegler A, et al. Early Prediction of Cardiac Complications in Acute Myocarditis by Means of Extracellular Volume Quantification With the Use of Dual-Energy Computed Tomography. JACC Cardiovasc Imaging 2021;14:2041-2. [Crossref] [PubMed]
6. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
7. Huang EP, Lin HS, Chen YC, et al. Lower attenuation and higher kurtosis of coronary artery calcification associated with vulnerable plaque - an agatston score propensity-matched CT radiomics study. BMC Cardiovasc Disord 2023;23:158. [Crossref] [PubMed]
8. Chen Q, Pan T, Wang YN, et al. A Coronary CT Angiography Radiomics Model to Identify Vulnerable Plaque and Predict Cardiovascular Events. Radiology 2023;307:e221693. [Crossref] [PubMed]
9. Jin X, Li Y, Yan F, et al. Automatic coronary plaque detection, classification, and stenosis grading using deep learning and radiomics on computed tomography angiography images: a multi-center multi-vendor study. Eur Radiol 2022;32:5276-86. [Crossref] [PubMed]
10. Si N, Shi K, Li N, et al. Identification of patients with acute myocardial infarction based on coronary CT angiography: the value of pericoronary adipose tissue radiomics. Eur Radiol 2022;32:6868-77. [Crossref] [PubMed]
11. Hinzpeter R, Wagner MW, Wurnig MC, et al. Texture analysis of acute myocardial infarction with CT: First experience study. PLoS One 2017;12:e0186876. [Crossref] [PubMed]
12. Mannil M, von Spiczak J, Manka R, et al. Texture Analysis and Machine Learning for Detecting Myocardial Infarction in Noncontrast Low-Dose Computed Tomography: Unveiling the Invisible. Invest Radiol 2018;53:338-43. [Crossref] [PubMed]
13. Antunes S, Esposito A, Palmisanov A, et al. Characterization of normal and scarred myocardium based on texture analysis of cardiac computed tomography images. Annu Int Conf IEEE Eng Med Biol Soc 2016;2016:4161-4. [Crossref] [PubMed]
14. Friedrich MG, Sechtem U, Schulz-Menger J, et al. Cardiovascular magnetic resonance in myocarditis: A JACC White Paper. J Am Coll Cardiol 2009;53:1475-87. [Crossref] [PubMed]
15. Ferreira VM, Schulz-Menger J, Holmvang G, et al. Cardiovascular Magnetic Resonance in Nonischemic Myocardial Inflammation: Expert Recommendations. J Am Coll Cardiol 2018;72:3158-76. [Crossref] [PubMed]
16. Kociol RD, Cooper LT, Fang JC, et al. Recognition and Initial Management of Fulminant Myocarditis: A Scientific Statement From the American Heart Association. Circulation 2020;141:e69-92. [Crossref] [PubMed]
17. Chinese Society of Cardiology, Chinese Medical Association, Writing Group. Chinese Society of Cardiology guidelines on the diagnosis and treatment of adult fulminant myocarditis. Sci China Life Sci 2024;67:913-39. [Crossref] [PubMed]
18. Bouleti C, Baudry G, Iung B, et al. Usefulness of Late Iodine Enhancement on Spectral CT in Acute Myocarditis. JACC Cardiovasc Imaging 2017;10:826-7. [Crossref] [PubMed]
19. Graça Santos L, Dias Ferreira N, Ladeiras-Lopes R, et al. Dual-energy cardiac computed tomography: the 'one-stop-shop' for acute myocarditis. Eur Heart J Cardiovasc Imaging 2021;22:129. [Crossref] [PubMed]
20. Zhuang B, Wang S, Zhao S, et al. Computed tomography angiography-derived fractional flow reserve (CT-FFR) for the detection of myocardial ischemia with invasive fractional flow reserve as reference: systematic review and meta-analysis. Eur Radiol 2020;30:712-25. [Crossref] [PubMed]
21. Zeng L, Yang Z, Qin P, et al. Dual Energy CT-based Radiomics For Identification Of Myocardial Focal Scar And Artificial Beam-hardening. J Cardiovasc Comput Tomogr 2024;18:S37-8. [Crossref]
22. Chen Y, Zhang N, Gao Y, et al. A coronary CT angiography-derived myocardial radiomics model for predicting adverse outcomes in chronic myocardial infarction. Int J Cardiol 2024;411:132265. [Crossref] [PubMed]
23. Lo Iacono F, Maragna R, Guglielmo M, et al. Identification of subclinical cardiac amyloidosis in aortic stenosis patients undergoing transaortic valve replacement using radiomic analysis of computed tomography myocardial texture. J Cardiovasc Comput Tomogr 2023;17:286-8. [Crossref] [PubMed]
24. Zhu MM, Zhu XM, Lin SS, et al. The incremental value of CCTA-derived myocardial radiomics signature for ischemia diagnosis with reference to CT myocardial perfusion imaging. Br J Radiol 2023;96:20220971. [Crossref] [PubMed]
25. Chou FS, Ghimire LV. Machine Learning for Mortality Prediction in Pediatric Myocarditis. Front Pediatr 2021;9:644922. [Crossref] [PubMed]
26. Baritussio A, Cheng CY, Lorenzoni G, et al. A Machine-Learning Model for the Prognostic Role of C-Reactive Protein in Myocarditis. J Clin Med 2022;11:7068. [Crossref] [PubMed]
27. Ferdeghini EM, Pinamonti B, Picano E, et al. Quantitative texture analysis in echocardiography: application to the diagnosis of myocarditis. J Clin Ultrasound 1991;19:263-70. [Crossref] [PubMed]
28. Lieback E, Meyer R, Romaniuk P, et al. Ultrasound sonographic diagnosis of myocarditis using quantitative texture analysis of 2-dimensional echocardiography images. Z Gesamte Inn Med 1989;44:484-7. [PubMed]
29. Papalini EI, Polte CL, Bobbio E, et al. Diagnosis of Acute Myocarditis Using Texture-Based Cardiac Magnetic Resonance, with CINE Imaging as a Novel Tissue Characterization Technique. Diagnostics (Basel) 2022;12:3187. [Crossref] [PubMed]
30. Cavallo AU, Di Donna C, Troisi J, et al. Radiomics analysis of short tau inversion recovery images in cardiac magnetic resonance for the prediction of late gadolinium enhancement in patients with acute myocarditis. Magn Reson Imaging 2022;94:168-73. [Crossref] [PubMed]
31. Di Noto T, von Spiczak J, Mannil M, et al. Radiomics for Distinguishing Myocardial Infarction from Myocarditis at Late Gadolinium Enhancement at MRI: Comparison with Subjective Visual Analysis. Radiol Cardiothorac Imaging 2019;1:e180026. [Crossref] [PubMed]
32. Baessler B, Luecke C, Lurz J, et al. Cardiac MRI Texture Analysis of T1 and T2 Maps in Patients with Infarctlike Acute Myocarditis. Radiology 2018;289:357-65. [Crossref] [PubMed]
33. Baessler B, Luecke C, Lurz J, et al. Cardiac MRI and Texture Analysis of Myocardial T1 and T2 Maps in Myocarditis with Acute versus Chronic Symptoms of Heart Failure. Radiology 2019;292:608-17. [Crossref] [PubMed]
34. Alis D, Yergin M, Asmakutlu O, et al. The influence of cardiac motion on radiomics features: radiomics features of non-enhanced CMR cine images greatly vary through the cardiac cycle. Eur Radiol 2021;31:2706-15. [Crossref] [PubMed]