Radiomics and machine learning applied to cine CMR images for the differential diagnosis of left ventricular hypertrophy etiologies

Bakground Accurate and non-invasive differentiation of cardiac amyloidosis, hypertrophic cardiomyopathy (HCM), and hypertensive heart disease remains a clinical challenge. While cardiac magnetic resonance (CMR) imaging with t1 mapping and late gadolinium enhancement (LGE) is the reference standard for myocardial characterization, its use may be limited by cost, scan duration, and contraindications in patients with renal impairment. Purpose This study aimed to assess the potential of radiomics and machine learning (ML) applied to CINE CMR sequences for the classification of these three conditions. Methods In this retrospective single-center study, we included 37 patients: 11 with cardiac amyloidosis, 8 with hypertensive heart disease, and 18 with HCM. Manual myocardial segmentation was performed at both systole and diastole. Radiomic features were extracted using PyRadiomics, generating five distinct datasets: systole, diastole, concatenated systole-diastole, the difference between diastole and systole (delta radiomics), and a combined dataset including all previous feature sets. Unsupervised feature selection based on correlation and mutual information was followed by a nested cross-validation pipeline using LASSO and Random Forest classifiers to optimize model performance and reproducibility. Results The most robust features identified were related to texture and morphological asymmetry, including skewness and surface-to-volume ratio (figure 1). The best-performing model, based on delta radiomics and diastolic image features, achieved a median accuracy of 86% (±11%). Repeated nested cross-validation and feature selection consistently identified a core subset of discriminative radiomic features, primarily reflecting myocardial texture and asymmetry. Model interpretability was supported using SHAP (SHapley Additive exPlanations) analysis. Conclusions Radiomics analysis of CINE CMR images, integrated with explainable machine learning, represents a promising, non-invasive approach for differentiating cardiac amyloidosis, HCM, and hypertensive heart disease. By relying solely on CINE sequences, it may reduce scan time, avoid contrast administration, and enhance accessibility. External validation in multicenter cohorts and integration of automated segmentation are warranted to enable clinical translation.Radiomic features