Beyond the visible: radiomic texture analysis of cardiovascular magnetic resonance for prediction of myocardial injury in acute myocarditis

Background Radiomic analysis of cardiovascular magnetic resonance (CMR) enables quantification of myocardial textural features that are not visible to the human eye with potential diagnostic and prognostic value in cardiovascular disease. The aim of this study was to predict myocardial injury, defined by late gadolinium enhancement (LGE), in acute myocarditis using artificial intelligence (AI) applied to clinical and CMR radiomic features. Methods The cohort included patients with a clinical diagnosis of acute myocarditis in a large tertiary centre. Left ventricular LGE extent was quantified using the 5SD threshold method and stratified into high (>10%) or low LGE (<10%) categories. Myocardial segmentation of short-axis cine and T1 mapping was performed using a trained nnUNet model (v2). Radiomic features were extracted using PyRadiomics (v3.1.0). Four feature sets were evaluated: clinical, T1 radiomics, clinical + T1 radiomics, and clinical + T1 + cine radiomics. Clinical variables included age, sex, time to CMR and troponin and cardiovascular risk factors such as hypertension, diabetes and smoking. Recursive feature elimination was utilised for feature selection. Supervised machine learning (ML) algorithms were evaluated across all feature sets including logistic regression, random forest and three boosting methods. Model performance was assessed using area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity. Feature importance was examined using SHAP (SHapley Additive exPlanations). Results 96 patients had short axis cine imaging, T1 mapping and LGE imaging with age 41.2 +/- 16.9 years. 61/96 patients were male and 67/96 were in the high LGE category. The best performing ML algorithms for each feature set are shown in Table 1. The AUROC to predict high LGE for the clinical + T1 feature set was 0.833 ± 0.069 (sensitivity 0.86, specificity 0.78) compared with the AUROC for the T1 feature set alone (0.830 ± 0.068 (sensitivity 0.90, specificity 0.79)). Texture-based radiomic descriptors derived from T1 mapping, predominantly from the grey-level co-occurrence matrix (GLCM) family appeared to be the most influential features for predicting high LGE (Figure 1). Conclusions CMR radiomics data can be leveraged to predict myocardial injury (LGE burden) in acute myocarditis. Adding clinical features to the T1 radiomic feature set results in minimal predictive improvement. Using an explainable AI approach, provides insight into why predictions are made and the mechanisms underlying myocardial injury in myocarditis. The most impactful radiomic features (GLCM), may reflect myocardial heterogeneity on T1 mapping and influence the prediction of the spatial extent of cardiac injury. Radiomics offers potential novel biomarkers for enhanced diagnosis. Predictive models could be used to guide disease stratification, and follow-up including the timing of repeat CMR study, and/or genetic testing.