Non-invasive diagnostic model for myocarditis using cardiac magnetic resonance radiomics

Introduction

Myocarditis is an inflammatory myocardial disease with a persistently high global incidence (10–22 per 100,000 population), contributing significantly to sudden cardiac death in young individuals and often progressing to heart failure (1,2). The clinical presentation is highly heterogeneous, ranging from asymptomatic cases to fatal outcomes, complicating early diagnosis (3-5). While endomyocardial biopsy (EMB) remains the diagnostic gold standard, its invasiveness and associated complications (e.g., a risk of cardiac tamponade of up to 9%) limit its use, particularly in early or non-severe cases (2,6). Non-invasive alternatives, such as serological markers (e.g., troponin) and electrocardiography, suffer from low specificity (<50%) and susceptibility to false positives under conditions like renal insufficiency or myocardial infarction (7-12). Echocardiography, though accessible, has limited sensitivity for detecting early myocardial edema and depends heavily on operator experience (13-15).

Cardiac magnetic resonance (CMR) has emerged as a cornerstone non-invasive tool for myocardial tissue characterization, leveraging multi-parameter imaging to comprehensively assess key pathological features of myocarditis, including edema, hyperemia, and fibrosis (16). The updated Lake Louise Criteria (LLC), endorsed by the Society for Cardiovascular Magnetic Resonance (SCMR), serve as the clinical reference standard for myocarditis diagnosis, achieving a pooled sensitivity of 74% and specificity of 86% (17-19). However, the full implementation of LLC relies heavily on advanced quantitative techniques such as T1/T2 mapping, which are classified as “advanced CMR applications” in SCMR guidelines (19). These techniques require high-field (3.0 T) magnetic resonance systems and specialized post-processing expertise, rendering them inaccessible in most primary care settings and resource-limited institutions globally (18,19).

Notably, conventional CMR sequences [cine sequence (CINE), T2-weighted imaging (T2WI), late gadolinium enhancement (LGE)] remain the most widely available modalities in clinical practice, as highlighted in SCMR’s standardized protocols for cardiovascular imaging (19). These sequences can effectively capture core pathological changes of early myocarditis (e.g., myocardial edema via T2WI and fibrotic scarring via LGE) without relying on mapping technology (16). Nevertheless, conventional CMR interpretation is predominantly qualitative and subject to substantial inter-observer variability, particularly among less experienced radiologists, which may lead to missed diagnoses of subtle early pathological alterations (17,19). This gap—between the inaccessibility of advanced mapping-based LLC and the interpretive limitations of widely available conventional sequences—underscores the need for optimized diagnostic strategies tailored to real-world clinical accessibility. Radiomics can extract high-throughput quantitative features from medical imaging and has shown potential in enhancing CMR-based diagnostics (20-22). Current radiomics studies applied to myocarditis report area under the curve (AUC) values ranging from 0.80 to 0.85; however, these studies are primarily limited to a single CMR sequence. For example, Baessler et al. utilized T1/T2 mapping radiomics to assess acute myocarditis (AUC =0.85), whereas Di Noto et al. focused on LGE-based features for differential diagnosis (AUC =0.82) (23,24). Common limitations of these models include single-modality design, which cannot capture all pathophysiological characteristics of myocarditis; lack of external validation, reducing generalizability; insufficient attention to feature harmonization across different imaging protocols, affecting reproducibility; and most studies lack transparent reporting and bias control.

Therefore, this study aims to construct and validate a machine learning model that combines multimodal CMR radiomics and clinical variables for the early, non-invasive diagnosis of myocarditis. As a foundational exploratory research, this study focuses on identifying the specific pathological features of acute myocarditis by distinguishing myocarditis patients from subjects with normal cardiac structure and function—this is a necessary prerequisite for subsequent model optimization to address the more complex clinical challenge of differentiating myocarditis from acute coronary syndrome, non-ischemic cardiomyopathies and other cardiovascular diseases with overlapping clinical and imaging features. The objective is to provide a tool with superior accuracy and generalizability, potentially reducing reliance on invasive procedures and supporting clinical decision-making in diverse healthcare settings, and to lay a foundation for subsequent clinical translational research on complex differential diagnosis. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2809/rc).

Methods

Patients

This study was approved by the Institutional Ethics Committee of Nanjing Drum Tower Hospital (No. 2024-551-01) and conducted in accordance with the principles outlined in the Declaration of Helsinki and its subsequent amendments. Renji Hospital was informed of and also approved this study. Informed consent was waived due to the retrospective design. Patients with myocarditis and those without myocarditis who visited two clinical centers from January 2017 to June 2025 were retrospectively enrolled. Clinical data and image data were extracted from the hospital’s electronic medical record Picture Archiving and Communication System (PACS) system, respectively.

Inclusion criteria for myocarditis group: (I) diagnosis based on the 2023 Japanese Circulation Society (JCS) criteria (13), which abandons the rigid requirement of myocyte necrosis for myocarditis diagnosis proposed by the Dallas Criteria (Aretz HT, 1987) (25) and regards inflammatory cell infiltration as the core histopathological evidence—this aligns with the classic viewpoint of Baughman KL (Circulation, 2006) (26) that myocyte necrosis is not an essential diagnostic index for myocarditis. The 2023 JCS criteria require histopathological evidence of inflammatory cell infiltration or, if EMB was not feasible, a combination of clinical symptoms (e.g., chest pain, elevated high-sensitivity troponin) and CMR findings consistent with myocardial edema; (II) availability of complete clinical records, including demographic data (age, gender), cardiovascular risk factors, and laboratory results; (III) high-quality CMR studies performed during the acute phase of illness (within 14 days of symptom onset), defined as images free of severe artifacts (e.g., motion-related distortions, signal dropouts) and with sufficient resolution for quantitative analysis; (IV) disease stage clarification: only patients with acute myocarditis (symptom duration <3 months) were included to minimize heterogeneity in radiomic feature extraction.

Inclusion criteria for non-myocarditis group: (I) absence of structural or functional cardiac abnormalities on CMR, as confirmed by two independent cardiologists; (II) full clinical documentation matching the scope of the myocarditis group; (III) CMR studies meeting the same technical quality standards, with explicit exclusion of individuals with histories of ischemic heart disease, cardiomyopathy, or other inflammatory conditions.

Exclusion criteria (applied to both groups): (I) incomplete clinical or imaging data (>5% missing variables); (II) suboptimal CMR image quality, such as sequences affected by significant motion artifacts, poor signal-to-noise ratio, or incomplete myocardial coverage; (III) comorbidities that could confound myocardial tissue characterization, including end-stage renal disease, hepatic failure, or active malignancies; (IV) chronic or subacute myocarditis (symptom duration >3 months) to ensure temporal consistency in radiomic analyses. The study workflow is shown in Figure 1.

Figure 1 Study workflow design. The red circles represent the delineated regions of interest (ROIs) for segmentation. CI, confidence interval; DCA, decision curve analysis; LOG, Laplacian of Gaussian; MAPSE, Mitral Annular Plane Systolic Excursion; MaxLAA, maximum left atrial area; MinLAV, minimum left atrial volume; OR, odds ratio; RAEF, right atrial ejection fraction; ROC, receiver operating characteristic curve; RVESV, right ventricular end-systolic volume.

Collection of clinical data

Based on the electronic medical record system, multi-dimensional health indicators of patients were collected, including basic demographic characteristics (gender, age), anthropometric parameters [body mass index (BMI)], lifestyle-related risk factors (smoking behavior, alcohol intake history), and structured data such as chronic disease diagnosis and treatment records (hypertension, diabetes mellitus history).

Image acquisition

This study used the Philips Medical 3.0T Achieva MRI system (Best, Netherlands). To minimize differences between sites, the same hardware configuration was maintained, including a dedicated 32-channel phased-array surface coil for cardiac imaging. The imaging protocol strictly followed the multi-sequence recommendations proposed by the SCMR (19), while also considering additional center-specific adjustments to ensure cross-institutional reproducibility. A standardized preoperative process was implemented for all subjects. Following ethical approval and informed consent, metal screening was performed to exclude contraindications. Participants then received guided breathing training to minimize motion artifacts. During imaging, subjects were positioned supine, and cardiac cycle synchronization was achieved using the Philips electrocardiographic vector synchronization system with retrospective electrocardiographic gating. The multi-sequence CMR protocol consisted of three core sequences: CINE, used to assess cardiac morphology and systolic function; T2WI (short tau inversion recovery), employed to detect myocardial edema (i.e., inflammatory exudate specific to myocardial inflammation, distinguished from passive interstitial edema caused by heart failure) associated with inflammation; and LGE (phase-sensitive inversion recovery) acquired 10–15 minutes post-contrast administration to identify fibrotic scars. Sequence parameters were optimized during pilot testing and fixed thereafter to ensure consistency. Key parameters—such as slice thickness, temporal resolution, and field of view—were harmonized between centers, with detailed specifications provided in Table S1.

CMR data measurement

In this study, the CVI42 professional analysis platform (version 6.1, Circle Cardiovascular Imaging) was used for standardized processing of CMR data, which was independently completed by two physicians with cardiovascular imaging diagnosis qualifications (Figure S1). (I) Ventricular function evaluation was based on continuous short-axis CINEs. By manually delineating the endocardial and epicardial boundaries of the ventricle at end-diastole and end-systole, core parameters such as ventricular ejection fraction (EF), stroke volume (SV), cardiac output (CO), cardiac index (CI), end-diastolic volume (EDV), and end-systolic volume (ESV) were quantitatively calculated. (II) Atrial function and longitudinal deformation analysis were performed using long-axis two-chamber and four-chamber cine images. Semi-automatic contour tracking technology was used to obtain the minimum/maximum atrial volume and atrial EF, and the mitral annular plane systolic excursion (MAPSE) and tricuspid annular plane systolic excursion (TAPSE) were simultaneously measured to evaluate atrial mechanical function comprehensively. (III) Myocardial strain analysis was performed by manually delineating the epicardial boundary of the right ventricle in end-diastolic short-axis and four-chamber views, combined with the global contour of the left ventricle (LV), to calculate the global circumferential strain (GCS), global longitudinal strain (GLS), and global radial strain (GRS) of the biventricle, systematically evaluating myocardial deformation capacity.

Image preprocessing and myocardial region of interest (ROI) delineation

Based on Python (version 3.11), this study integrates the SimpleITK, NumPy, and scikit-image toolkits to construct an automated image registration workflow. The specific preprocessing steps are as follows: First, cardiac image data in DICOM format, including short-axis CINEs, T2WI, and LGE, were imported using the three-dimensional (3D) Slicer platform (version 5.61). Previous research has focused chiefly on the end-diastolic phase of the myocardium, as the myocardial boundaries are more stable at this phase, so end-diastolic data were selected (27-29). The data were converted to NIfTI format and subjected to multi-level preprocessing—using the N4ITK algorithm to correct low-frequency intensity artifacts caused by MRI field inhomogeneities and iteratively optimizing to eliminate grayscale non-uniformity; an adaptive normalization method based on Otsu’s threshold was used to map pixel values to the (0, 1) range, eliminating differences in grayscale across different devices.

Spatial registration used the CINE as the fixed reference image, adopting a two-stage registration strategy: initial rigid registration calculated geometric center alignment parameters through CenteredTransformInitializer; subsequent affine optimization based on Mattes mutual information metric (50 histogram bins) was performed, configuring a gradient descent optimizer (learning rate 1.0, maximum iterations 300, convergence threshold 1e−6) and B-spline interpolation algorithm to achieve sub-pixel-level spatial alignment. The registered images were resampled to the reference coordinate system using the Resample function, and a high-quality, multimodal, aligned dataset was then evaluated using the structural similarity index and output.

A physician with five years of experience in cardiac imaging diagnosis performed the delineation of the myocardial ROI. Specifically, the epicardial and endocardial boundaries of the LV were manually drawn to segment the left ventricular myocardium. Additionally, the papillary muscles were excluded to prevent non-myocardial tissue structures from interfering with the radiomic features. Meanwhile, to evaluate the intra-observer and inter-observer consistency of radiomic features, another radiologist with the same qualifications was invited to participate in the delineation, and both radiologists were blinded to the patients’ clinical information (Figure 1).

Feature extraction and selection

Using Python’s PyRadiomics toolkit (version 3.0), 1,130 radiomic features were extracted from each unimodal image (CINE, T2WI, LGE) in accordance with the Image Biomarker Standardization Initiative (IBSI) guidelines, comprehensively covering the morphological and texture characteristics of myocardial tissue. All images were first resampled to an isotropic voxel spacing of 1 mm × 1 mm × 1 mm using B-spline interpolation to ensure spatial consistency. Grayscale discretization was performed using the fixed bin width method with a binwidth of 25. Specifically, the extracted features included three categories based on original images: shape features (e.g., geometric parameters such as volume and surface area), first-order statistical features (e.g., metrics reflecting grayscale distribution including mean and variance), and texture feature combinations (e.g., gray-level co-occurrence matrix capturing pixel spatial correlation); in addition, transformed-image based features were included, comprising Laplacian of Gaussian (LoG) features [processed with Gaussian filtering and Laplacian operator, sigma values = (1, 2, 3)] and wavelet features (processed with wavelet transform using the Daubechies 4 wavelet family as the base function). Feature extraction was performed on the 3D volumetric ROI formed by concatenating consecutive myocardial slices, rather than individual two-dimensional (2D) slices, to capture spatial correlations across the entire left ventricular myocardium. The above transformation features can extract fine structural information, complementing the original image features and thereby more comprehensively depict the pathophysiological characteristics of myocardial tissue.

During the machine learning process, the central dataset was first randomly divided into a training set (for feature selection and model training) and a validation set (for model validation) at a ratio of 7:3. First, for the data in the training set, radiomic feature selection was performed as follows: (I) intraclass correlation coefficient (ICC) analysis was first conducted to assess the reproducibility of radiomic features, and only features with an ICC value greater than 0.8 were retained to ensure high stability and reliability; (II) Mann-Whitney test (P<0.05) was used to screen for features with significant intergroup differences, completing preliminary feature selection; (III) Pearson correlation analysis (calculating the absolute value of the correlation coefficient between features) was used to eliminate highly correlated features (correlation coefficient threshold |r|>0.8) to eliminate multicollinearity between features; (IV) least absolute shrinkage and selection operator (LASSO) regression combined with 5-fold cross-validation was used for further selection (100 alpha values were used for LASSO path calculation, and the maximum number of iterations for model training was 1,000), retaining features with an average absolute coefficient greater than 0 in cross-validation and ranking them according to importance. Selection of clinical predictors: first, univariate logistic regression analysis was used to initially screen potential predictors, setting a significance threshold of P<0.10 to minimize the omission of practical features; then, a multivariate Logistic Regression (LR) model was used for further selection, retaining features with statistical significance (P<0.05).

Model construction

To identify the optimal machine learning algorithm for constructing the myocarditis diagnostic model, a comprehensive evaluation of multiple classifiers was conducted. Seven widely used algorithms were implemented and compared, including k-Nearest Neighbors (KNNs), Decision Tree (DT), LR, Support Vector Classifier (SVC), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Random Forest (RF). The performance of each classifier was systematically evaluated using 5-fold cross-validation on the training set based on the area under the receiver operating characteristic (ROC) curve (AUC). The algorithm demonstrating the highest average AUC across cross-validation folds was selected as the final classifier for subsequent model development. For the integration of multimodal data, a pre-fusion strategy at the feature level was employed, wherein radiomic features extracted from different CMR sequences (CINE, T2WI, LGE) were concatenated with clinical parameters to form a combined feature vector, which was then input into the selected classifier. This approach ensured comprehensive utilization of complementary information from different data modalities while maintaining model interpretability.

Model evaluation

The model evaluation was conducted using a comprehensive, multidimensional metric system to rigorously assess the performance and generalizability of the constructed predictors. The optimal classifier identified through comparative analysis was employed to train the model on the training dataset, followed by performance evaluation on the internal validation set and generalization verification on an independent external test set. The area under the ROC curve with 95% confidence intervals (CI) served as the primary metric for quantifying diagnostic efficacy. At the same time, the Brier score complemented the calibration curves in assessing prediction accuracy. Additional evaluation components included decision curve analysis (DCA) for clinical net benefit assessment across probability thresholds, as well as standard classification metrics (accuracy, precision, and specificity) for comprehensive performance characterization. This multi-faceted approach ensured robust evaluation of discrimination capability, calibration performance, and clinical utility.

Statistical methods

R software (Version 4.3.2) was used for analysis. During the preprocessing stage, variables with missing values exceeding 30% were excluded, and the remaining missing values were imputed using the mean of continuous variables and the mode of categorical variables. The Shapiro-Wilk test was used to evaluate data normality. Normally distributed data were expressed as mean ± standard deviation, and non-normally distributed data were described as median and interquartile range (Q1–Q3).

Results

Clinical features

In this retrospective study, a total of 724 patients were included, and 378 were excluded, resulting in 346 patients enrolled (Figure 2), including 185 patients with myocarditis (53.78%). Patients from Center 1 (Nanjing Drum Tower Hospital) were randomly divided into a training set (n=197) and an internal validation set (n=85) at a ratio of 7:3. Center 2 (Shanghai Renji Hospital) provided an external validation set (n=64). The proportions of myocarditis patients in the three cohorts were 50.8%, 50.6%, and 50.0%, respectively (Table 1). In the training set, significant differences were observed in age, hypertension incidence, right ventricular end-systolic volume (RVESV), mitral annular plane systolic excursion (MAPSE) lateral, and right atrial ejection fraction (RAEF) between the myocarditis group and the non-myocarditis group (P<0.05). Still, no differences in hypertension incidence and RAEF were found in the internal validation set (P>0.05). In contrast, age, RVESV, and RAEF also showed significant differences in the external validation set (P<0.05) (Table 1). Univariate logistic regression analysis of the training set showed that 7 clinical variables were retained. Multivariate logistic regression analysis revealed that age [odds ratio (OR) =0.976, 95% CI: 0.956–0.996, P=0.019] and MAPSE Lateral (OR =0.979, 95% CI: 0.883–1.087, P=0.009) were significantly associated with the risk of myocarditis (Table 2).

Figure 2 Flowchart showing the patient selection process from two institutions. CMR, cardiac magnetic resonance.

Radiomic feature selection

In this study, 1,130 radiomic features were extracted from the myocardial regions of the CINE, T2WI sequence, and LGE sequence, respectively. For the training set, Mann-Whitney test (P<0.05) was used to screen 290 CINE, 187 T2, and 151 LGE features with significant intergroup differences, completing preliminary feature selection; then, Pearson correlation analysis was used to eliminate 245 highly correlated features in the CINE, 164 in T2, and 131 in LGE (correlation coefficient threshold |r|>0.8) to eliminate multicollinearity between features; in addition, LASSO regression combined with 5-fold cross-validation was used for further feature selection, finally retaining 12 CINE, 11 T2, and 10 LGE non-zero coefficient features (Figure S2). Based on the features selected from each unimodal above, the fusion features of the imaging sequences were further selected, and a total of 12 radiomic features were retained based on the three imaging sequences (Figure 3). For the Fusion_Clinical_Radiomic model (which integrates radiomics and clinical features), a comprehensive feature importance analysis was further conducted based on an RF classifier to identify key diagnostic features. Ultimately, 14 most valuable diagnostic features for acute myocarditis were confirmed—12 CMR radiomic features from three sequences and 2 clinically significant variables (age and MAPSE Lateral) (Figure 4). These features collectively capture the multidimensional pathological changes of acute myocarditis.

Figure 3 Imaging features selected by different models. (A) Radiomic fusion model. (B) Clinical-radiomic fusion model. 3D, three-dimensional; CINE, cine sequence; HLH, high-pass low-pass high-pass; HLL, high-pass low-pass low-pass; LGE, late gadolinium enhancement; LHL, low-pass high-pass low-pass; LLH, low-pass low-pass high-pass; LLL, low-pass low-pass low-pass; MAPSE, Mitral Annular Plane Systolic Excursion.

Figure 4 SHAP-based feature interpretability analysis of the Fusion_Clininc_Radiomic model. (A) SHAP feature importance bar chart. (B) SHAP feature contribution dot plot. 3D, three-dimensional; CINE, cine sequence; HLL, high-pass low-pass low-pass; LGE, late gadolinium enhancement; LHH, low-pass high-pass high-pass; LHL, low-pass high-pass low-pass; LLH, low-pass low-pass high-pass; LLL, low-pass low-pass low-pass; MAPSE, Mitral Annular Plane Systolic Excursion; SHAP, Shapley Additive exPlanation.

Model construction and evaluation

To identify the optimal machine learning algorithm for model construction, a comprehensive comparative analysis of seven classifiers was conducted, including KNN, DT, LR, SVC, XGBoost, LightGBM, and RF. These classifiers were initially evaluated on the training set and internal validation set using core performance metrics. The RF algorithm demonstrated balanced and superior performance across multiple evaluation dimensions. Specifically, in the internal validation set, it achieved an AUC of 0.827 and an accuracy of 0.776 under the clinical-radiomic feature-level fusion modality, representing the optimal performance among all classifiers. More importantly, RF exhibited stable performance across unimodal (CINE, T2WI, LGE) and multimodal fusion scenarios without significant overfitting (e.g., KNN achieved perfect scores on the training set but demonstrated poor generalization on the validation set) or metric instability (e.g., DT showed low specificity in certain modalities). Its generalization capability and classification stability were significantly superior to other algorithms, as detailed in Table S2.

Based on the selected features, an RF classifier was employed as the base model, and hyperparameter tuning was performed through a combination of grid search and 5-fold cross-validation. Three categories of predictive models were ultimately constructed: (I) unimodal models, which included one clinical model (Clinical) and three radiomic models (CINE, T2WI, LGE). The AUC values for these unimodal models in the training set ranged from 0.726 (95% CI: 0.656–0.796) to 0.789 (95% CI: 0.725–0.853), with accuracies between 0.660 and 0.736. Their Brier scores varied from 0.209 to 0.256, indicating moderate calibration performance. In the external validation set, the AUC values ranged from 0.648 (95% CI: 0.510–0.787) to 0.738 (95% CI: 0.615–0.862), with specificities between 0.281 and 0.906. The corresponding Brier scores in the external set (0.227–0.252) showed a slight increase in prediction error compared to the training set. (II) The radiomic fusion model (Fusion_Radiomic) demonstrated improved performance, achieving an AUC of 0.893 (95% CI: 0.849–0.937), an accuracy of 0.817, a specificity of 0.825, and a Brier score of 0.199 in the training set. In the external validation set, it maintained an AUC of 0.810 (95% CI: 0.697–0.922) and an accuracy of 0.813, with a Brier score of 0.225, reflecting its superior performance over the unimodal models. (III) The clinical-radiomic multimodal fusion model (Fusion_Clinical_Radiomic), which integrated clinical features with radiomic features from three CMR sequences via a pre-fusion strategy, exhibited the most robust performance. It achieved an AUC of 0.936 (95% CI: 0.905–0.967), an accuracy of 0.848, a specificity of 0.845, and the most favorable Brier score of 0.141 in the training set. In the external validation set, this model maintained a high AUC of 0.854 (95% CI: 0.754–0.953) and a specificity of 0.906, with a Brier score of 0.193, demonstrating significantly better performance than both the radiomic fusion model and all unimodal models across all datasets (Table 3).

Figure 5 (ROC curve) shows the discrimination performance of the 6 models. The AUC values of the Fusion_Clinical_Radiomic model in the training set, internal validation set, and external validation set were 0.936 (95% CI: 0.905–0.967), 0.862 (95% CI: 0.779–0.944), and 0.854 (95% CI: 0.754–0.953), respectively, and it performed best in the internal validation set. Detailed performance indicators of each model are shown in Table 3. The DeLong test results of AUC between models (Table S3) showed that: the AUC of the Fusion_Radiomic model was statistically different from all unimodal models in the training set and external validation set (P<0.05), and only not statistically different from the CINE model in the internal validation set (P=0.231); while the AUC of the Fusion_Clinical_Radiomic model was statistically different from all other models in the training set, internal validation set, and external validation set (P<0.05) (Figure 6). Figure 7 (DCA curve) suggests that the multimodal fusion model has a wide range of positive net benefit threshold probability intervals in the three cohorts, indicating good clinical application potential; in addition, the Hosmer-Lemeshow test (all P>0.05) combined with 5,000 bootstrap resamples demonstrated that the calibration curves of the model in each dataset were closely aligned with the ideal diagonal, indicating reliable calibration performance without significant departure from perfect calibration (Figure 8).

Figure 5 ROC curves of different models. (A) Training set. (B) Internal validation set. (C) External validation set. AUC, area under the curve; CINE, cine sequence; FPR, false positive rate; LGE, late gadolinium enhancement; ROC, receiver operating characteristic curve; TPR, true positive rate.

Figure 6 Performance comparison of different models in the three cohorts using the DeLong test. (A) Training set. (B) Internal validation set. (C) External validation set. **, P<0.01; ***, P<0.001. CINE, cine sequence; LGE, late gadolinium enhancement.

Figure 7 DCA curves of different models. (A) Training set. (B) Internal validation set. (C) External validation set. CINE, cine sequence; DCA, decision curve analysis; LGE, late gadolinium enhancement.

Figure 8 Calibration curves of Fusion_Radiomics model and Fusion_Clinical_Radiomics model. (A) Training set. (B) Internal validation set. (C) External validation set.

Discussion

This study developed and validated a multimodal machine learning model integrating radiomic features from three CMR sequences (CINE, T2WI, LGE) with clinical parameters for the early, non-invasive diagnosis of myocarditis. The Fusion_Clinical_Radiomic model demonstrated superior and stable performance, achieving AUC values of 0.936 (95% CI: 0.905–0.967), 0.862 (95% CI: 0.779–0.944), and 0.854 (95% CI: 0.754–0.953) in the training, internal validation, and external validation sets, respectively. The model’s performance was consistently superior to unimodal approaches and radiomic-only fusion models across all datasets, underscoring the value of integrating multimodal data.

Previous radiomics studies in myocarditis have primarily relied on single CMR sequences, limiting their ability to capture the full spectrum of disease pathophysiology. For example, the updated LLC incorporate T1 and T2 mapping techniques, achieving a pooled sensitivity of 78% and specificity of 88% (AUC =0.83) in meta-analyses, but they often lack integration with functional sequences or clinical parameters (30). Baessler et al. (23) conducted texture analysis of T1 and T2 Mapping to distinguish infarct-like acute myocarditis in 2018 (AUC =0.85), but did not include the CINE for evaluating cardiac function and only targeted patients with specific subtypes, resulting in limited generalization; another study by Baessler et al. in 2019 (31) distinguishing acute from chronic myocarditis (AUC =0.83) also failed to break through the limitation of unimodality; Di Noto et al. (24) used LGE radiomics to distinguish myocardial infarction from myocarditis in 2019 (AUC =0.82), with the core goal of “differential diagnosis” rather than “early diagnosis”. In contrast, our multimodal approach synergistically integrates: T2WI—to quantify inflammatory edema via hyperintense signals, LGE—to identify fibrotic scars through LGE, CINE—to detect subtle motion abnormalities indicative of systolic dysfunction. This comprehensive coverage of pathological stages—from edema (T2WI) to fibrosis (LGE) and functional impairment (CINE)—explains the higher AUC of our fusion model (0.854) compared to unimodal studies (AUC 0.80–0.85). Furthermore, while Chen et al. combined CTA radiomics with clinical features, their model incurred radiation exposure and used a smaller sample (32). Our study avoids these limitations by leveraging non-invasive CMR and multicenter data.

The enhanced performance of the Fusion_Clinical_Radiomic model stems from its alignment with the pathobiology of myocarditis. Myocardial edema (inflammatory exudate from increased vascular permeability in myocardial inflammation), an early hallmark of inflammation, is captured by T2WI radiomic features reflecting water-content changes and tissue heterogeneity (33). Fibrotic remodeling, a consequence of chronic injury, is quantified via LGE-based texture analysis (34). Meanwhile, CINE-derived strain parameters detect subclinical systolic dysfunction often preceding overt structural damage (35). The inclusion of clinical variables further refines risk stratification: age correlates with disease prevalence trends (peak incidence: 30–45 years) (36), while MAPSE serves as an early marker of longitudinal dysfunction independent of load conditions (37). This multiparametric integration enables the model to identify complementary pathological processes, reducing diagnostic gaps inherent in single-modality approaches.

The observed sensitivity drop in external validation (internal: 0.767 vs. external: 0.531) primarily stems from site heterogeneity, including variations in scanner protocols, population demographics (e.g., age distribution), and disease-stage prevalence across centers. These factors may compromise feature stability and generalizability. Key limitations include the retrospective design, which introduces potential confounding biases despite multicenter validation; reliance on manual ROI delineation, which is time-consuming and subjective despite high segmentation consistency; dependency on single-vendor data (Philips 3.0 T), limiting cross-platform applicability; and the absence of subgroup analyses for etiologies (e.g., viral, autoimmune) or disease stages (acute/chronic), restricting subtype-specific insights, and an age imbalance between the myocarditis and non-myocarditis groups (myocarditis patients are significantly younger). This age imbalance reflects the recognized epidemiology of acute myocarditis (peak incidence in young and middle-aged adults). To address these issues, first, future work should prioritize prospective multicenter trials with standardized imaging protocols to minimize heterogeneity; second, integrating deep learning for automated segmentation could enhance efficiency and objectivity; third, expanding datasets to include multi-vendor and multi-field-strength (e.g., 1.5 T/3.0 T) CMR images would improve cross-equipment robustness; fourth, we will expand the study cohort to include patients with acute coronary syndrome, non-ischemic cardiomyopathies and other cardiovascular diseases with overlapping clinical/imaging features with myocarditis, to optimize the model and improve its ability to solve complex real-world clinical differential diagnosis problems; fifth, including a larger subgroup of patients with EMB confirmation in future prospective studies to enable validation against the histological gold standard, which was not feasible in the current retrospective cohort due to scarce accessible data; and finally, establishing larger cohorts to enable stratified analyses by etiology and disease stage, which is essential to validate and refine the model for precise clinical application.

Conclusions

In conclusion, based on retrospective data from two centers, this study integrated radiomic features from conventional CMR sequences (CINE, T2WI, LGE) with clinical parameters to construct and validate the multimodal fusion Fusion_Clinical_Radiomic model for the early non-invasive diagnosis of myocarditis. The model exhibited stable performance in the external validation set, with an AUC of 0.854 (95% CI: 0.754–0.953), excellent calibration, and substantial clinical net benefit. Its diagnostic performance was superior to that of unimodal models during the same period. Future studies should further validate the model through prospective multicenter trials, incorporate data from multi-vendor equipment, optimize automatic segmentation algorithms, and conduct stratified analyses by etiological subtypes and disease courses to enhance the model’s generalizability and clinical translational value.

Acknowledgments

The researchers express their gratitude to Shanghai Tenth People’s Hospital, Nanjing Drum Tower Hospital, and Renji Hospital.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2809/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2809/dss

Funding: This work was supported by the National Natural Science Foundation of China General Program (No. 82272065); Shanghai Municipal Health Commission 2025 Smart Healthcare Special Project (No. 2025ZHYL020); Shanghai Science and Technology Commission 2025 Basic Research Program “Explorer Plan” (First Batch) (No. 25TS1406100); Shanghai Shenkang Hospital Development Center Diagnosis and Treatment Technology Promotion and Optimization Management Project, Municipal Hospital Emerging Frontier Joint Research Project (No. SHDC12026110); 2024 Nanjing Drum Tower Hospital Clinical Research Special General Project (No. 2024-LCYJ-MS-12); and Tongji University Medicine-X Interdisciplinary Research Initiative (No. 2025-0553-YB-11).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2809/coif). All authors report the funding from the National Natural Science Foundation of China General Program (No. 82272065); Shanghai Municipal Health Commission 2025 Smart Healthcare Special Project (No. 2025ZHYL020); Shanghai Science and Technology Commission 2025 Basic Research Program “Explorer Plan” (First Batch) (No. 25TS1406100); Shanghai Shenkang Hospital Development Center Diagnosis and Treatment Technology Promotion and Optimization Management Project, Municipal Hospital Emerging Frontier Joint Research Project (No. SHDC12026110); 2024 Nanjing Drum Tower Hospital Clinical Research Special General Project (No. 2024-LCYJ-MS-12); and Tongji University Medicine-X Interdisciplinary Research Initiative (No. 2025-0553-YB-11). Xiuzheng Yue and Xiao Yu are employees of Philips Healthineers. The authors have no other 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 Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Ethics Committee of Nanjing Drum Tower Hospital (No. 2024-551-01), and Renji Hospital was informed of and also approved this study. Informed consent was waived due to the retrospective design.

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. Global, regional, and national incidence, prevalence, and years lived with disability for 301 acute and chronic diseases and injuries in 188 countries, 1990-2013: a systematic analysis for the Global Burden of Disease Study 2013. Lancet 2015;386:743-800.
2. Lampejo T, Durkin SM, Bhatt N, Guttmann O. Acute myocarditis: aetiology, diagnosis and management. Clin Med (Lond) 2021;21:e505-10. [Crossref] [PubMed]
3. Sagar S, Liu PP, Cooper LT Jr. Myocarditis. Lancet 2012;379:738-47. [Crossref] [PubMed]
4. Grossman SM, Pravda NS, Orvin K, Hamdan A, Vaturi M, Bengal T, Kornowski R, Weissler-Snir A. Characterization and long-term outcomes of patients with myocarditis: a retrospective observational study. Postepy Kardiol Interwencyjnej 2021;17:60-7. [Crossref] [PubMed]
5. Lynge TH, Nielsen TS, Gregers Winkel B, Tfelt-Hansen J, Banner J. Sudden cardiac death caused by myocarditis in persons aged 1-49 years: a nationwide study of 14 294 deaths in Denmark. Forensic Sci Res 2019;4:247-56. [Crossref] [PubMed]
6. Singh V, Mendirichaga R, Savani GT, Rodriguez A, Blumer V, Elmariah S, Inglessis-Azuaje I, Palacios I. Comparison of Utilization Trends, Indications, and Complications of Endomyocardial Biopsy in Native Versus Donor Hearts (from the Nationwide Inpatient Sample 2002 to 2014). Am J Cardiol 2018;121:356-63. [Crossref] [PubMed]
7. Al-Biltagi M, Issa M, Hagar HA, Abdel-Hafez M, Aziz NA. Circulating cardiac troponins levels and cardiac dysfunction in children with acute and fulminant viral myocarditis. Acta Paediatr 2010;99:1510-6. [Crossref] [PubMed]
8. Korff S, Katus HA, Giannitsis E. Differential diagnosis of elevated troponins. Heart 2006;92:987-93. [Crossref] [PubMed]
9. Chen JY, Lee SY, Li YH, Lin CY, Shieh MD, Ciou DS. Urine High-Sensitivity Troponin I Predict Incident Cardiovascular Events in Patients with Diabetes Mellitus. J Clin Med 2020;9:3917. [Crossref] [PubMed]
10. Shah KS, Maisel AS, Fonarow GC. Troponin in Heart Failure. Heart Fail Clin 2018;14:57-64. [Crossref] [PubMed]
11. Younis A, Matetzky S, Mulla W, Masalha E, Afel Y, Chernomordik F, Fardman A, Goitein O, Ben-Zekry S, Peled Y, Grupper A, Beigel R. Epidemiology Characteristics and Outcome of Patients With Clinically Diagnosed Acute Myocarditis. Am J Med 2020;133:492-9. [Crossref] [PubMed]
12. Buttà C, Zappia L, Laterra G, Roberto M. Diagnostic and prognostic role of electrocardiogram in acute myocarditis: A comprehensive review. Ann Noninvasive Electrocardiol 2020;25:e12726. [Crossref] [PubMed]
13. Nagai T, Inomata T, Kohno T, Sato T, Tada A, Kubo T, et al. JCS 2023 Guideline on the Diagnosis and Treatment of Myocarditis. Circ J 2023;87:674-754. [Crossref] [PubMed]
14. Løgstrup BB, Nielsen JM, Kim WY, Poulsen SH. Myocardial oedema in acute myocarditis detected by echocardiographic 2D myocardial deformation analysis. Eur Heart J Cardiovasc Imaging 2016;17:1018-26. [Crossref] [PubMed]
15. Hsiao JF, Koshino Y, Bonnichsen CR, Yu Y, Miller FA Jr, Pellikka PA, Cooper LT Jr, Villarraga HR. Speckle tracking echocardiography in acute myocarditis. Int J Cardiovasc Imaging 2013;29:275-84. [Crossref] [PubMed]
16. Ammirati E, Frigerio M, Adler ED, Basso C, Birnie DH, Brambatti M, Friedrich MG, Klingel K, Lehtonen J, Moslehi JJ, Pedrotti P, Rimoldi OE, Schultheiss HP, Tschöpe C, Cooper LT Jr, Camici PG. Management of Acute Myocarditis and Chronic Inflammatory Cardiomyopathy: An Expert Consensus Document. Circ Heart Fail 2020;13:e007405. [Crossref] [PubMed]
17. Ferreira VM, Schulz-Menger J, Holmvang G, Kramer CM, Carbone I, Sechtem U, Kindermann I, Gutberlet M, Cooper LT, Liu P, Friedrich MG. Cardiovascular Magnetic Resonance in Nonischemic Myocardial Inflammation: Expert Recommendations. J Am Coll Cardiol 2018;72:3158-76. [Crossref] [PubMed]
18. Liu A, Wijesurendra RS, Francis JM, Robson MD, Neubauer S, Piechnik SK, Ferreira VM. Adenosine Stress and Rest T1 Mapping Can Differentiate Between Ischemic, Infarcted, Remote, and Normal Myocardium Without the Need for Gadolinium Contrast Agents. JACC Cardiovasc Imaging 2016;9:27-36. [Crossref] [PubMed]
19. Kramer CM, Barkhausen J, Bucciarelli-Ducci C, Flamm SD, Kim RJ, Nagel E. Standardized cardiovascular magnetic resonance imaging (CMR) protocols: 2020 update. J Cardiovasc Magn Reson 2020;22:17. [Crossref] [PubMed]
20. Elmahdy M, Sebro R. Radiomics analysis in medical imaging research. J Med Radiat Sci 2023;70:3-7. [Crossref] [PubMed]
21. Leiner T. Radiomics for Predicting Risk of Sudden Cardiac Death in Hypertrophic Cardiomyopathy. JACC Cardiovasc Imaging 2024;17:28-30. [Crossref] [PubMed]
22. Jordan MI, Mitchell TM. Machine learning: Trends, perspectives, and prospects. Science 2015;349:255-60. [Crossref] [PubMed]
23. Baessler B, Luecke C, Lurz J, Klingel K, von Roeder M, de Waha S, Besler C, Maintz D, Gutberlet M, Thiele H, Lurz P. Cardiac MRI Texture Analysis of T1 and T2 Maps in Patients with Infarctlike Acute Myocarditis. Radiology 2018;289:357-65. [Crossref] [PubMed]
24. Di Noto T, von Spiczak J, Mannil M, Gantert E, Soda P, Manka R, Alkadhi H. 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]
25. Aretz HT. Myocarditis: the Dallas criteria. Hum Pathol 1987;18:619-24. [Crossref] [PubMed]
26. Baughman KL. Diagnosis of myocarditis: death of Dallas criteria. Circulation 2006;113:593-5. [Crossref] [PubMed]
27. Wang S, Ling H, Yu J, He W, Li X, Wang Y, Huang L, Zheng J, Chen Y, Peng L. Association of diastolic dysfunction with myocardial tissue characteristics assessed by multi-parameter cardiac magnetic resonance in patients with idiopathic inflammatory myopathy. Rheumatology (Oxford) 2026;65:keaf535. [Crossref] [PubMed]
28. 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]
29. Burešová M, Pavlíček J, Hanzlíková P, Tomášková H, Rybníček O. 2D speckle tracking echocardiography and comparison with cardiac magnetic resonance in children with acute myocarditis. Front Cardiovasc Med 2024;11:1446602. [Crossref] [PubMed]
30. Kotanidis CP, Bazmpani MA, Haidich AB, Karvounis C, Antoniades C, Karamitsos TD. Diagnostic Accuracy of Cardiovascular Magnetic Resonance in Acute Myocarditis: A Systematic Review and Meta-Analysis. JACC Cardiovasc Imaging 2018;11:1583-90. [Crossref] [PubMed]
31. Baessler B, Luecke C, Lurz J, Klingel K, Das A, von Roeder M, de Waha-Thiele S, Besler C, Rommel KP, Maintz D, Gutberlet M, Thiele H, Lurz P. 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]
32. Chen X, Lv L, Pan J, Guan D, Huang Y, Hu Y, Zhang H, Hu H. Development of a clinical prediction model for acute myocarditis using coronary computed tomography angiography-based radiomics. Cardiovasc Diagn Ther 2025;15:85-99. [Crossref] [PubMed]
33. Gräni C, Eichhorn C, Bière L, Murthy VL, Agarwal V, Kaneko K, Cuddy S, Aghayev A, Steigner M, Blankstein R, Jerosch-Herold M, Kwong RY. Prognostic Value of Cardiac Magnetic Resonance Tissue Characterization in Risk Stratifying Patients With Suspected Myocarditis. J Am Coll Cardiol 2017;70:1964-76. [Crossref] [PubMed]
34. Yang F, Wang J, Li W, Xu Y, Wan K, Zeng R, Chen Y. The prognostic value of late gadolinium enhancement in myocarditis and clinically suspected myocarditis: systematic review and meta-analysis. Eur Radiol 2020;30:2616-26. [Crossref] [PubMed]
35. Fischer K, Obrist SJ, Erne SA, Stark AW, Marggraf M, Kaneko K, Guensch DP, Huber AT, Greulich S, Aghayev A, Steigner M, Blankstein R, Kwong RY, Gräni C. Feature Tracking Myocardial Strain Incrementally Improves Prognostication in Myocarditis Beyond Traditional CMR Imaging Features. JACC Cardiovasc Imaging 2020;13:1891-901. [Crossref] [PubMed]
36. Spotts PH, Zhou F. Myocarditis and Pericarditis. Prim Care 2024;51:111-24. [Crossref] [PubMed]
37. Cirin L, Crișan S, Luca CT, Buzaș R, Lighezan DF, Văcărescu C, Cozgarea A, Tudoran C, Cozma D. Mitral Annular Plane Systolic Excursion (MAPSE): A Review of a Simple and Forgotten Parameter for Assessing Left Ventricle Function. J Clin Med 2024;13:5265. [Crossref] [PubMed]