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

Background: Myocarditis is an inflammatory myocardial disease with high clinical heterogeneity, and its early diagnosis relies on invasive endomyocardial biopsy (EMB) or advanced cardiac magnetic resonance (CMR) techniques with poor accessibility in primary hospitals. Conventional CMR interpretation is subjective with inter-observer variability, and single-modality radiomics models have insufficient diagnostic accuracy. This study aimed to construct and validate a machine learning model integrating multimodal CMR radiomics features and clinical parameters for early non-invasive and accurate diagnosis of myocarditis.

Methods: This study recruited 344 participants from two hospitals (185 with myocarditis and 159 without myocarditis). Radiomics features were extracted from the left ventricular myocardium using multimodal CMR sequences [cine sequence (CINE), T2-weighted imaging (T2WI), and late gadolinium enhancement imaging (LGE)] and combined with clinical parameters. Feature selection was performed sequentially: first, intraclass correlation coefficient (ICC) analysis was used to retain features with ICC >0.8 for reproducibility; then, Mann-Whitney U test, Pearson correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression (for dimensionality reduction). For clinical predictors, univariate logistic regression (P<0.10) was used for preliminary screening, and multivariate logistic regression (P<0.05) for further selection. Models were constructed using random forests, including single-modality, radiomics fusion, and clinical-radiomics fusion variants. Model performance was evaluated in the training set, internal validation set, and external validation set.

Results: The Fusion_Clinical_Radiomic model exhibited optimal and stable diagnostic performance, with area under the curve (AUC) values of 0.936 [95% confidence interval (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 set, internal validation set, and external validation set, respectively. Its accuracy, specificity, and sensitivity in the training set were 0.848, 0.845, and 0.850, respectively, with a favorable Brier score of 0.141. In the external validation set, it maintained a specificity of 0.906 and a Brier score of 0.193. The model’s performance was significantly superior to all unimodal models and the radiomics fusion model alone (all P<0.05). Decision curve analysis (DCA) showed that the model had extensive clinical net benefits.

Conclusions: The fusion model provides an accurate, non-invasive auxiliary diagnostic tool for the diagnosis of acute myocarditis within 14 days of symptom onset. By relying on conventional CMR sequences (CINE, T2WI, LGE), it may reduce the reliance on invasive EMB and enhance diagnostic accessibility in primary hospitals. Supported by multicenter internal and external validation, the model exhibits favorable generalizability, offering a promising reference for clinical practice in the early identification of acute myocarditis.