Early assessment of myocardial injury in patients with coronavirus disease 2019 using a two-stage deep learning framework based on non-contrast chest computed tomography

Introduction

The pandemic outbreak of novel coronavirus disease 2019 (COVID-19) ensured that infection and spread are inevitable, but its prognosis is of great concern (1,2). The disease affects not only the respiratory system but also the cardiovascular system, with myocardial injury (MI) playing a key role in progression, prognosis, and treatment (3). The incidence of acute MI in patients with COVID-19 is high, up to 30%, and it causes 40% of deaths, especially among hospitalized patients (4,5). Troponin elevation among patients with COVID-19 increases the risk of mortality compared with patients without MI, which is also commonly used in the clinical diagnosis of MI (6,7), but most often for patients presenting with symptoms of acute coronary syndrome (ACS) (8). Consequently, MI diagnosis is frequently delayed in asymptomatic patients.

Non-contrast chest computed tomography (CT) is routinely performed in COVID-19 patients to assess pulmonary lesions (9). If non-contrast chest CT can simultaneously detect lung lesions and MI in patients with COVID-19 at the first admission examination, it is undoubtedly a good choice for the early screening of MI and simplifying the next step of the diagnosis and treatment process, especially for elderly and critically ill patients. However, it is well known that radiologists cannot simply use chest CT to identify whether patients with COVID-19 have MI (10,11). The emergence of artificial intelligence (AI) has increased the scope of diagnosis in traditional imaging (12). There have been some attempts to distinguish between normal and scar myocardial tissue in myocarditis and myocardial infarction with texture analysis method of cardiac CT plain scan (13,14). However, these studies had small sample sizes and did not conduct meaningful receiver operating characteristic (ROC) curve analysis on the test dataset, and the texture analysis was limited to a single image of interest. Deep learning (DL) is a subfield of machine learning (15), which can learn abstract representations and patterns of data in layered architectures of neural networks, achieving fusion of anatomy and function, improving the diagnostic efficiency of clinicians and conserving medical resources (16-19). Given that myocarditis and myocardial infarction are the relevant pathological manifestations of MI associated with COVID-19 (20,21), non-contrast chest CT based on DL is expected to serve as a new method for detecting myocardial lesions, providing the possibility to expand the clinical application of CT.

Currently, there have been few corresponding reports for COVID-19 patients. Hence, the purpose of this study was to develop a DL framework to effectively predict the MI of patients with COVID-19 in a one-step manner based on non-contrast chest CT, achieving early screening of MI when admission or primary diagnosis of COVID-19 and reducing the trouble and financial burden of those elderly and critically ill patients. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-449/rc).

Methods

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The Affiliated Hospital of Southwest Medical University (No. KY2023279) and the requirement for individual consent for this analysis was waived due to the retrospective nature. All data were anonymized to protect patient privacy.

Patients

From December 2022 to February 2023, 540 consecutive patients with laboratory-confirmed COVID-19 who underwent non-contrast chest CT were retrospectively recruited, comprising 280 patients with MI and 260 patients without MI. COVID-19-related MI was defined as a high-sensitivity troponin T (hsTnT) level above the 99th-percentile upper reference limit (0.14 ng/mL) (22,23). Cardiac troponin T (hsTnT below 0.14 ng/mL), and negative results from the electrocardiogram (ECG) and echocardiography were regarded as patients without MI. The inclusion criteria were as follows: (I) troponin levels assessed between 72 hours before and 48 hours after the COVID-19 diagnosis, according to the Fourth Universal Definition of Myocardial Infarction (24); (II) CT examination time was within 3 days after the first troponin measurement; and (III) age was at least 18 years old. The exclusion criteria were as follows (25): (I) known cardiomyopathy, previous myocarditis, and cardiac surgery; (II) previous coronary artery disease (evidence of coronary artery stenosis >50%) or previous myocardial infarction; (III) moderate to severe valve dysfunction; (IV) previous heart failure or atrial fibrillation; (V) uncontrolled hypertension; and (VI) poor image quality with artifacts. Finally, 230 COVID-19 patients with MI [aged 72.9±14.1 years (range, 18–98 years), 129 males] and 223 COVID-19 patients without MI [aged 57.8±16.9 years (range, 18–87 years), 105 males] were recruited to this study.

We collected baseline demographics, lifestyle factors (smoking/alcohol use), comorbidities [hypertension, diabetes, chronic obstructive pulmonary disease (COPD), liver disease, malignancy, obesity], clinical parameters (symptoms, vital signs, COVID-19 severity), and diagnostic results (laboratory tests, ECGs, echocardiography). The workflow of this study is illustrated as Figure 1.

Figure 1 The flowchart of this study. (A) The entire dataset was randomly divided into the training test and testing test. (B) The first DL module automatically identified the ROI of LV, and the second module performed binary classification to distinguish between MI and non-MI cases based on the extracted ROI. (C) The two modules were independently tested. ACC, accuracy; AUC, area under the curve; CT, computed tomography; DL, deep learning; GT, ground truth; IoU, intersection over union; LV, left ventricle; MI, myocardial injury; ROC, receiver operating characteristic; ROI, region of interest.

Non-contrast CT protocol

All patients were scanned in a supine position, arms up, head first, with a scanning range from the apex to the bottom of the lung, using GE Optima CT540 16 slice CT scanner (GE HealthCare, Chicago, IL, USA) or dual-layer spectral-detector CT (IQon Spectral CT, Philips Healthcare, Best, The Netherlands). To minimize motion artifacts, patients were instructed on breath-holding. The scans were acquired and reconstructed as axial images using the following parameters: 5.0 mm section thickness, 5.0 mm interval, 120 kVp, with adaptive tube current.

Manual data annotation

The DL utilized in this study was supervised models, therefore we invited a team of radiologists with at least 5 years of cardiac CT diagnosis experience to manually annotate non-contrast chest CT images. For each image, experts were asked to delineate the contour of left ventricle (LV) as region of interest (ROI) using LabelMe (5.0.1), a publicly available annotation tool in DL studies. The left anterior descending artery (LAD) is used as the boundary between the left and right ventricles, while the left circumflex artery (LCX) is used as the boundary between the LV and the left atrium (LA). LV range was defined from the planes of the mitral valve to the apex. The annotation data were extracted and used for subsequent DL training and evaluations. Positive and negative cases of MI were determined according to the criteria for MI and reviewed by the team. For all data, the team reached consensus to ensure the annotation quality.

Two-stage DL framework

In this study, we proposed a two-stage DL framework. The first DL module automatically identified the ROI of LV to obtain the segmentations. The second DL module took the segmentations obtained in the first stage and conducted a binary classification of positives and negatives of MI. We evaluated several state-of-the-art DL architectures and found that FCN-ResNet-101 (26) and DenseNet-121 (27) performed best, therefore we mainly introduced these two models and reported their corresponding results in this study. FCN-ResNet-101 provided pixel-wise segmentation with enhanced deep feature representation through residual connections, which is particularly beneficial in delineating complex anatomical structures. DenseNet-121, with its dense connectivity, promotes feature reuse and alleviates vanishing gradients, making it highly effective for medical image classification. Namely, the first module was based on the DL architecture of FCN-ResNet-101, as illustrated in Figure 2A. FCN-ResNet-101 is a deep neural network combined of fully-connected network (FCN) and backbone of residual network (ResNet). FCN-ResNet-101 takes advantages of both structures, making it especially good at sematic segmentations. Taking non-contrast images as inputs, FCN-ResNet101 generates masks of the predicted ROIs of LV in the same sizes of inputs. The second module was based on DenseNet-121, as illustrated in Figure 2B. DenseNet-121 is a deep neural network using the dense block as the main structure. It concentrates on features and makes full use of all features in the training process by densely connecting all previous layers with the following layers, thus achieving better results and reducing the number of parameters. In DenseNet-121, the dense connection method was designed to make full use of features. More specifically, the features of each layer are saved and used in each subsequent convolution operation. Therefore, all the previous information (before and after the convolution operation) is used in each layer of the network architecture. Based on the obtained segmentations of ROIs of LV obtained by FCN-ResNet-101 in the first stage, DenseNet-121 extracts abstract representations of ROIs and output a tuple of two possibilities indicating the likeliness of positive MI or negative MI. Combining the two DL modules, the proposed DL framework could automatically identify the LV region, and based on which, further classify the status of MI as positive or negative.

Figure 2 The architecture of segmentation and classification modules. (A) Segmentation module FCN-ResNet-101 was used to identify the left ventricle. (B) Classification module DenseNet-121 was used to distinguish MI and non-MI. FCN, fully-connected network; MI, myocardial injury; ResNet, residual network; ReLU, rectified linear unit.

In the preparation of datasets, we first randomly split the data at the patient level into a training set and a testing set. The training set included 413 patients with 3,837 images, whereas 40 patients with 152 images comprised the testing set. There was no patient overlap between the two sets. The training dataset was used to train the two modules, namely FCN-ResNet-101 for segmentations and DenseNet-121 for binary classifications. All images were preprocessed before training, including normalization and resizing to 320×320 pixels. In training, the two DL modules were trained separately as independent tasks. We adopted a four-fold cross-validation approach in the training-validation of two tasks to obtain the best-performing trained models, which were tested later in the independent testing subsets. In this framework, the models were developed using an epoch of 30, an optimizer of Adam, a learning rate of 0.0001, and a batch size of one.

The data processing and the DL framework were fully implemented in Python (3.9.7). Several common Python language packages were utilized including Pandas (1.3.4), cv2 (4.8.0) and Pillow (8.4.0) for data handling, PyTorch (1.11.0) for training of DL modules, and matplotlib (3.2.2) for data visualizations. The DL framework was trained in a specialized DL training server equipped with advanced graphic processing unit (GPU), NVIDIA RTX 3090 (NVIDIA, Santa Clara, CA, USA) with 24 GB graphic memory, 62 GB main memory, and Intel^® Xeon^® 2.10 GHz Intel central processing unit (CPU). The training took 0.2 hours for one epoch. However, the trained DL framework could inference within acceptable time in a typical desktop computer with general settings, making it more accessible in clinical practices.

Statistical analysis

The statistical software SPSS 26.0 (IBM Corp., Armonk, NY, USA) was used for statistical analysis. Continuous variables were expressed as mean ± standard deviation/median range. The categorical variables were summarized as frequencies, with the χ² test used to test differences between two groups. A two-tailed P value of less than 0.05 was deemed statistically significant.

The segmentation model outputs probability maps of the LV region, whereas the classification model generates probabilities of MI. Both outputs are binarized using a threshold of 0.5 to produce the final segmentation masks and diagnostic classifications, respectively. In the performance evaluation of the first segmentation DL module, we utilized metric of intersection over union (IoU) and Dice coefficient. A larger value of IoU indicated a good overlap of the predicted segmentation with the ground truth of ROIs. For the binary classification of the second DL module, we plotted the ROC and calculated the area under the curve (AUC) with 95% confidence interval (CI). We also reported the value of sensitivity (SEN), specificity (SPE), accuracy (ACC), and F1 score. We plotted the classification results as confusion matrices and ROC diagrams.

Results

Patient characteristics

The clinical characteristics of 453 patients are reported in Table 1. There was no significant difference in gender and body mass index (BMI) between patients with and without MI (P>0.05), but there was a significant difference in age (P<0.05). The patients with MI were generally older than those without MI. The MI group showed significantly higher cardiac troponin levels (0.086±0.324 µg/L) compared to controls (0.007±0.003 µg/L). Although fever and cough rates were similar at admission, respiratory distress was more prevalent in the MI group. These patients also had higher rates of critical COVID-19 (P<0.001) and lower rates of non-critical disease (P<0.001). Notably, the MI group had a higher prevalence of hypertension (45.2% vs. 21.2%), diabetes (25.4% vs. 13.4%), COPD, and ECG abnormalities (24.8% vs. 17.0%). No significant differences were observed in cancer, liver disease, smoking, or alcohol consumption between groups.

Performance of DL framework

Segmentation of LV

In the first stage of our proposed DL framework, FCN-ResNet101 automatically generated the masks of segmentations of LV. First, in the cross-validation of training using the dataset, the mean IoU was 0.8079. We tested FCN-ResNet-101 in the testing dataset, and the model obtained an IoU of 0.8041, an ACC of 0.9949, and a Dice score of 0.8672. We visualized the segmentations in some randomly selected samples in Figure 3. As displayed, the DL module could accurately identify the ROIs of LV in non-contrast CT images.

Figure 3 Segmentation results of the FCN-ResNet-101. The original CT image is shown for visual reference. The ground truth (blue) represents the left ventricle region manually annotated by experts, while the prediction (red) indicates the segmentation result generated by the model. CT, computed tomography; FCN, fully-connected network; MI, myocardial injury; ResNet, residual network.

Classification of patients with MI and non-MI

In the second stage of our DL framework, DenseNet-121 conducted the classification to discriminate the status of MI. On the testing dataset, the DL module achieved an AUC of 0.8618 (95% CI: 0.8049–0.9187), an SPE of 0.6071, a SEN of 0.8750, an ACC of 0.7763, and an F1-score of 0.8317. We summarized the results in Table 2 and visualized the confusion matrices and diagrams of ROC for the three validation datasets in Figure 4. The results showed that the DL could accurately determine the MI statuses with encouraging performance.

Figure 4 Classification results of the DenseNet-121. (A) Confusion matrix. (B) Receiver operating characteristic curve. AUC, area under the curve; MI, myocardial injury.

Discussion

Main findings

Our DL framework enables MI detection in COVID-19 patients using routine non-contrast chest CT, providing a practical one-stop solution for combined pulmonary and cardiac assessment. This one-stop approach could enable early detection, allowing high-risk patients to undergo further tests [e.g., echocardiography or cardiac magnetic resonance imaging (MRI)] for detailed evaluation-optimizing medical resources. Non-contrast chest CT can identify MI prior to clinical troponin detection, providing a novel and practical approach for early risk stratification in COVID-19 patients.

The feasibility of AI for helping to diagnose MI based on non-contrast CT

Acute MI can range from asymptomatic elevation of cardiac troponins to fulminant myocarditis and circulatory shock in COVID-19 patients (4,28). Research shows that patients with elevated troponin are at a higher risk for adverse events during hospitalization, including a higher death rate, acute respiratory distress syndrome, and malignant arrhythmias (2,29). After adjusting for disease severity and related clinical factors, even a small amount of MI is significantly associated with death (23). Therefore, for asymptomatic patients with MI, it is often easy to overlook the cardiovascular system in clinical practice, which may affect the progression of the patient’s disease. Reports of MI in association with COVID-19 have included acute ischemic injury (type 1 myocardial infarction) as well as nonischemic injury (i.e., myocarditis), stress cardiomyopathy, acute heart failure, and secondary cardiac injury caused by sepsis and critical illness (20,21,30,31). Cardiovascular magnetic resonance (CMR) is the reference standard for the assessment of acute MI because it can assess structure, function, scarring, and inflammation (32). A research report on COVID-19 using CMR showed that myocarditis-like injury pattern was common (33,34), and there was myocardial infarction (35). Recent autopsy studies suggest that in COVID-19-related MI, microvascular thrombosis leading to microinfarction is more frequent than myocarditis (36-38). COVID-19 patients with MI demonstrate significantly more cardiac abnormalities, particularly ventricular systolic dysfunction, pericardial effusion, and myocardial scarring. These scars mainly stem from myocardial infarction and microinfarction, emphasizing the thrombotic sequelae of COVID-19. These early microscopic pathological changes will inevitably lead to changes in myocardial tissue density, spatial distribution of internal structure, and other high-dimensional features. With the help of DL, it can break through the limitations of the human eye, fully explore useful information in images, and provide the possibility for us to screen for MI on non-contrast chest CT. In our study, the construction of the DL-based model used 3,989 CT images (230 cases with MI and 223 cases without MI). The DL-based model yielded an AUC of 0.8618, with high ACC, SEN, precision, and F1-score. With the aid of the model, it is potential and feasible for radiologists to diagnose COVID-19 and screen MI early based on non-contrast chest CT. Compared with other examinations, it is more suitable for the early screening of risk stratification of patients with COVID-19.

LV segmentation on non-contrast chest CT

Cardiac chamber segmentations are generally performed on cardiac computed tomography angiography (CCTA) scans, due to the presence of contrast material that provides accurate identification of the different cardiac chambers. However, during the COVID-19 pandemic, COVID-19 infected patients often only received non-contrast chest CT for various reasons (detection or screening). Left ventricular function is an important predictor of outcomes for patients with myocardial dysfunction. Accurate segmentation of the LV is crucial for our research, as reproducible data is needed to solve the problem of manual segmentation. The cardiac landmarks are a set of point landmarks associated to cardiac structures, such as chamber centers, valve centers, or vessel ostia (39). The positions of cardiac landmark sets are essential reference structure for the segmenting of LV. In addition, the center of the LV (a simple shape with some symmetry properties) is a well-defined landmark. This provides an anatomical basis for our cardiac imaging doctors to manually label the LV using non-contrast chest CT. In our first step of segmentation, our proposed DL framework accurately obtained the segmentation of the LV in non-contrast chest CT, with an IoU of 0.8041, an ACC of 0.9949, and a Dice coefficient of 0.8672.

Considerations of applying AI in MI diagnosis

First, AI methods, especially DL, rely on adequate high-quality data. Larger datasets from more patients with diverse conditions will lead to more accurate and robust DL models (40). However, in this pilot single-center study, the data size was limited. Fortunately, the proposed DL framework is ready to be extended for future multicenter study. Second, considering the small size of LV in the whole image, in the present two-stage framework, we utilized DL models to identify the LV before classifying MI. In DL-based medical image analysis, previous studies have conducted classifications using the whole image. In these studies, the ROI usually took a considerable portion of the image, allowing a single-stage DL framework to be possible. Our experiments showed that the two-stage framework is significantly superior to direct classification using the whole image rather than the ROI. We believe that this two-stage framework is not only suitable for MI diagnosis but also has important implications for other scenarios of DL-based diagnosis. Third, the present DL framework only utilized CT information. In multimodal DL, we could fuse CT with other modalities such as clinical features. In future work, it is possible to extend the classification DL model in the second stage to fuse more information for possible better performance. In addition, this study considered combinations of representative segmentation models and classification models. However, as DL models advance at a fast rate, it would be necessary to investigate new state-of-the-art models in hope of further improving the accuracies.

MI and comorbidities

Consistent with previous findings (2,5), COVID-19 patients with MI were typically older and more likely to present with critical illness than those without MI. These patients also demonstrated higher prevalence rates of hypertension, diabetes, and COPD. These findings underscore the importance of early screening and prompt diagnosis of MI in those high-risk populations to improve clinical outcomes. Although the COVID-19 pandemic has ended, the sequelae following infection still warrant attention and research. Studies have shown that COVID-19 patients with myocarditis have more severe myocardial edema during short-term follow-up than those without myocarditis (41). In the future, we plan to contact individuals who have recovered from COVID-19 to further investigate whether MI persists.

Limitations

Our study has some limitations. First, most of the patients we included were tested for cardiac troponin due to symptoms suspected of acute MI. The data of asymptomatic or chronic MI patients were not analyzed in this study. Moving forward, we will collect more patients with other types of MI in COVID-19 to improve this model. Second, our patients did not undergo gold standard tests such as CMR or myocardial radionuclides because they were hospitalized patients with severe conditions. However, our research is based on clinical and previous literature to obtain this part of data, which can be further studied based on CMR or other criteria in the future.

Conclusions

The proposed two-stage DL framework could fully automatedly identify LV in non-contrast chest CT images of COVID-19 patients and dragonize the MI statuses with promising performance, demonstrating feasibility and potentials in clinical applications.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-449/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-449/dss

Funding: This study was supported by the General Project of National Natural Science Foundation (No. 82272077), Key Research and Development Project of Science & Technology Department of Sichuan Province (No. 2020YFS0324), General Project of Natural Science of Chongqing Science and Technology Bureau (No. CSTB2022NSCQ-MSX0854), and the Research Project of Southwest Medical University (No. 2023XGZX007).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-449/coif). The 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. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The Affiliated Hospital of Southwest Medical University (No. KY2023279) and individual consent for this analysis was waived due to the retrospective nature.

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. Ruan Q, Yang K, Wang W, Jiang L, Song J. Clinical predictors of mortality due to COVID-19 based on an analysis of data of 150 patients from Wuhan, China. Intensive Care Med 2020;46:846-8. Erratum in: Intensive Care Med 2020;46:1294-7. [Crossref] [PubMed]
2. Guo T, Fan Y, Chen M, Wu X, Zhang L, He T, Wang H, Wan J, Wang X, Lu Z. Cardiovascular Implications of Fatal Outcomes of Patients With Coronavirus Disease 2019 (COVID-19). JAMA Cardiol 2020;5:811-8. [Crossref] [PubMed]
3. Sandoval Y, Januzzi JL Jr, Jaffe AS. Cardiac Troponin for Assessment of Myocardial Injury in COVID-19: JACC Review Topic of the Week. J Am Coll Cardiol 2020;76:1244-58. [Crossref] [PubMed]
4. Bavishi C, Bonow RO, Trivedi V, Abbott JD, Messerli FH, Bhatt DL. Special Article - Acute myocardial injury in patients hospitalized with COVID-19 infection: A review. Prog Cardiovasc Dis 2020;63:682-9. [Crossref] [PubMed]
5. Shi S, Qin M, Shen B, Cai Y, Liu T, Yang F, Gong W, Liu X, Liang J, Zhao Q, Huang H, Yang B, Huang C. Association of Cardiac Injury With Mortality in Hospitalized Patients With COVID-19 in Wuhan, China. JAMA Cardiol 2020;5:802-10. [Crossref] [PubMed]
6. Lippi G, Lavie CJ, Sanchis-Gomar F. Cardiac troponin I in patients with coronavirus disease 2019 (COVID-19): Evidence from a meta-analysis. Prog Cardiovasc Dis 2020;63:390-1. [Crossref] [PubMed]
7. Imazio M, Klingel K, Kindermann I, Brucato A, De Rosa FG, Adler Y, De Ferrari GM. COVID-19 pandemic and troponin: indirect myocardial injury, myocardial inflammation or myocarditis? Heart 2020;106:1127-31. [Crossref] [PubMed]
8. Amsterdam EA, Wenger NK, Brindis RG, Casey DE Jr, Ganiats TG, Holmes DR Jr, Jaffe AS, Jneid H, Kelly RF, Kontos MC, Levine GN, Liebson PR, Mukherjee D, Peterson ED, Sabatine MS, Smalling RW, Zieman SJ. 2014 AHA/ACC Guideline for the Management of Patients with Non-ST-Elevation Acute Coronary Syndromes: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines. J Am Coll Cardiol 2014;64:e139-228. Erratum in: J Am Coll Cardiol 2014;64:2713-4. [Crossref] [PubMed]
9. Xie X, Zhong Z, Zhao W, Zheng C, Wang F, Liu J. Chest CT for Typical Coronavirus Disease 2019 (COVID-19) Pneumonia: Relationship to Negative RT-PCR Testing. Radiology 2020;296:E41-5. [Crossref] [PubMed]
10. Zhong Y, Sun Z, Xu P, Bai Y, Zhang Z, Wang G. The value of non-contrast chest CT in the prediction of myocardial injury in patients with the COVID-19 Omicron variant. Sci Rep 2023;13:10321. [Crossref] [PubMed]
11. Li S, Wang X, Hu H, Xu J, He J, Yang W, et al. Additional Value of Non-contrast Chest CT in the Prediction of Adverse Cardiovascular Events in Patients With Novel Coronavirus Disease 2019 (COVID-19). Front Cardiovasc Med 2021;8:738044. [Crossref] [PubMed]
12. Johnson KW, Torres Soto J, Glicksberg BS, Shameer K, Miotto R, Ali M, Ashley E, Dudley JT. Artificial Intelligence in Cardiology. J Am Coll Cardiol 2018;71:2668-79. [Crossref] [PubMed]
13. Antunes S, Esposito A, Palmisanov A, Colantoni C, de Cobelli F, Del Maschio A. 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. Mannil M, von Spiczak J, Manka R, Alkadhi H. 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]
15. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436-44. [Crossref] [PubMed]
16. Chow R, Midroni J, Kaur J, Boldt G, Liu G, Eng L, Liu FF, Haibe-Kains B, Lock M, Raman S. Use of artificial intelligence for cancer clinical trial enrollment: a systematic review and meta-analysis. J Natl Cancer Inst 2023;115:365-74. [Crossref] [PubMed]
17. Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL. Artificial intelligence in radiology. Nat Rev Cancer 2018;18:500-10. [Crossref] [PubMed]
18. Lüscher TF, Wenzl FA, D’Ascenzo F, Friedman PA, Antoniades C. Artificial intelligence in cardiovascular medicine: clinical applications. Eur Heart J 2024;45:4291-304. [Crossref] [PubMed]
19. Niazi MKK, Parwani AV, Gurcan MN. Digital pathology and artificial intelligence. Lancet Oncol 2019;20:e253-61. [Crossref] [PubMed]
20. Doyen D, Moceri P, Ducreux D, Dellamonica J. Myocarditis in a patient with COVID-19: a cause of raised troponin and ECG changes. Lancet 2020;395:1516. [Crossref] [PubMed]
21. Deng Q, Hu B, Zhang Y, Wang H, Zhou X, Hu W, Cheng Y, Yan J, Ping H, Zhou Q. Suspected myocardial injury in patients with COVID-19: Evidence from front-line clinical observation in Wuhan, China. Int J Cardiol 2020;311:116-21. [Crossref] [PubMed]
22. Puelacher C, Lurati Buse G, Seeberger D, Sazgary L, Marbot S, Lampart A, et al. Perioperative Myocardial Injury After Noncardiac Surgery: Incidence, Mortality, and Characterization. Circulation 2018;137:1221-32. [Crossref] [PubMed]
23. Lala A, Johnson KW, Januzzi JL, Russak AJ, Paranjpe I, Richter F, et al. Prevalence and Impact of Myocardial Injury in Patients Hospitalized With COVID-19 Infection. J Am Coll Cardiol 2020;76:533-46. [Crossref] [PubMed]
24. Kini A, Cao D, Nardin M, Sartori S, Zhang Z, Pivato CA, Chiarito M, Nicolas J, Vengrenyuk Y, Krishnamoorthy P, Sharma SK, Dangas G, Fuster V, Mehran R. Types of myocardial injury and mid-term outcomes in patients with COVID-19. Eur Heart J Qual Care Clin Outcomes 2021;7:438-46. [Crossref] [PubMed]
25. Wang H, Li R, Zhou Z, Jiang H, Yan Z, Tao X, Li H, Xu L. Cardiac involvement in COVID-19 patients: mid-term follow up by cardiovascular magnetic resonance. J Cardiovasc Magn Reson 2021;23:14. [Crossref] [PubMed]
26. Shelhamer E, Long J, Darrell T. Fully Convolutional Networks for Semantic Segmentation. IEEE Trans Pattern Anal Mach Intell 2017;39:640-51. [Crossref] [PubMed]
27. Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected convolutional networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI, USA: IEEE; 2017.
28. Zeng JH, Liu YX, Yuan J, Wang FX, Wu WB, Li JX, Wang LF, Gao H, Wang Y, Dong CF, Li YJ, Xie XJ, Feng C, Liu L. First case of COVID-19 complicated with fulminant myocarditis: a case report and insights. Infection 2020;48:773-7. [Crossref] [PubMed]
29. Zhou F, Yu T, Du R, Fan G, Liu Y, Liu Z, Xiang J, Wang Y, Song B, Gu X, Guan L, Wei Y, Li H, Wu X, Xu J, Tu S, Zhang Y, Chen H, Cao B. Clinical course and risk factors for mortality of adult inpatients with COVID-19 in Wuhan, China: a retrospective cohort study. Lancet 2020;395:1054-62. [Crossref] [PubMed]
30. Chitsazan M, Amin A, Chitsazan M, Ziaie N, Amri Maleh P, Pouraliakbar H, Von Haehling S. Heart failure with preserved ejection fraction in coronavirus disease 2019 patients: the promising role of diuretic therapy in critically ill patients. ESC Heart Fail 2021;8:1610-4. [Crossref] [PubMed]
31. Jirak P, Larbig R, Shomanova Z, Fröb EJ, Dankl D, Torgersen C, Frank N, Mahringer M, Butkiene D, Haake H, Salzer HJF, Tschoellitsch T, Lichtenauer M, Egle A, Lamprecht B, Reinecke H, Hoppe UC, Pistulli R, Motloch LJ. Myocardial injury in severe COVID-19 is similar to pneumonias of other origin: results from a multicentre study. ESC Heart Fail 2021;8:37-46. [Crossref] [PubMed]
32. Leiner T, Bogaert J, Friedrich MG, Mohiaddin R, Muthurangu V, Myerson S, Powell AJ, Raman SV, Pennell DJ. SCMR Position Paper (2020) on clinical indications for cardiovascular magnetic resonance. J Cardiovasc Magn Reson 2020;22:76. [Crossref] [PubMed]
33. Nagel E, Puntmann VO. Errors in Statistical Numbers and Data in Study of Cardiovascular Magnetic Resonance Imaging in Patients Recently Recovered From COVID-19. JAMA Cardiol 2020;5:1307-8. [Crossref] [PubMed]
34. Puntmann VO, Carerj ML, Wieters I, Fahim M, Arendt C, Hoffmann J, Shchendrygina A, Escher F, Vasa-Nicotera M, Zeiher AM, Vehreschild M, Nagel E. Outcomes of Cardiovascular Magnetic Resonance Imaging in Patients Recently Recovered From Coronavirus Disease 2019 (COVID-19). JAMA Cardiol 2020;5:1265-73. [Crossref] [PubMed]
35. Kotecha T, Knight DS, Razvi Y, Kumar K, Vimalesvaran K, Thornton G, et al. Patterns of myocardial injury in recovered troponin-positive COVID-19 patients assessed by cardiovascular magnetic resonance. Eur Heart J 2021;42:1866-78. [Crossref] [PubMed]
36. Artico J, Shiwani H, Moon JC, Gorecka M, McCann GP, Roditi G, et al. Myocardial Involvement After Hospitalization for COVID-19 Complicated by Troponin Elevation: A Prospective, Multicenter, Observational Study. Circulation 2023;147:364-74. [Crossref] [PubMed]
37. Kawakami R, Sakamoto A, Kawai K, Gianatti A, Pellegrini D, Nasr A, Kutys B, Guo L, Cornelissen A, Mori M, Sato Y, Pescetelli I, Brivio M, Romero M, Guagliumi G, Virmani R, Finn AV. Pathological Evidence for SARS-CoV-2 as a Cause of Myocarditis: JACC Review Topic of the Week. J Am Coll Cardiol 2021;77:314-25. [Crossref] [PubMed]
38. Clinical presentation, disease course, and outcome of COVID-19 in hospitalized patients with and without pre-existing cardiac disease: a cohort study across 18 countries. Eur Heart J 2022;43:1104-20. [Crossref] [PubMed]
39. Lorenz C, von Berg J. A comprehensive shape model of the heart. Med Image Anal 2006;10:657-70. [Crossref] [PubMed]
40. Lenharo M. An AI revolution is brewing in medicine. What will it look like? Nature 2023;622:686-8.
41. Chen X, Hu Y, Pan J, Ye L, Pan Y, Liu Q. Multiparametric cardiovascular magnetic resonance in patients with myocarditis with consecutive follow-up and a comparison between non-COVID-19 and COVID-19-associated myocarditis. Quant Imaging Med Surg 2025;15:486-501. [Crossref] [PubMed]