Cardiovascular magnetic resonance radiomics for phenotyping myocardial fibrosis in severe aortic stenosis

Background Myocardial fibrosis (MF) is a key determinant of adverse outcomes in severe aortic stenosis and limits the efficacy of valve implantation. Endomyocardial biopsy is the reference standard for diagnosing MF, but its invasiveness limits routine use. Cardiovascular magnetic resonance (CMR)-derived extracellular volume fraction (ECV%) is a non-invasive surrogate but correlates only modestly with biopsy-proven MF. Purpose We investigated whether comprehensive radiomic analysis of CMR images improves the non-invasive identification of increased biopsy-proven MF burden. Methods We prospectively enrolled 56 patients with severe aortic stenosis scheduled for transcatheter aortic valve implantation. All patients underwent pre-procedural 3T CMR and intraprocedural endomyocardial biopsy. Histologic MF was quantified by digital morphometry, with patients classified as high (≥11% collagen area fraction) or low fibrotic burden. CMR mid-ventricular short-axis myocardium was manually segmented on cine, native T1 maps, and late gadolinium enhancement images. Radiomics features were extracted using the PyRadiomics package. Only features with intra-/interobserver correlation ≥0.80 were included, and outcome-blinded principal component analysis was applied for dimensionality reduction. Five machine learning models, each using five-fold cross-validation, were tested to model the association between radiomics PCs and MF. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), compared with an ECV%-only model and a combined ECV% + radiomics PCs model. Radiomics PCs were correlated with clinical and imaging features to contextualise the radiomics signature (Figure 1). Results Median age was 80 [77-83] years, and 30% were female. A total of 29 patients (52%) had a high MF burden. Patients with higher MF had larger LV end-diastolic volumes (92 vs 76mL/m²; p=0.018) and showed a trend towards higher ECV% (28% vs 26%; p=0.072). For detecting an increased MF burden, the ECV%-only model achieved an AUC of 0.62 [0.57–0.67]. The machine learning models based on five cine radiomics PCs outperformed ECV% alone (AUC 0.80 [0.77-0.83] for logistic regression, similar for other models). Adding ECV% to the radiomics model did not provide additional discriminatory benefit (ΔAUC −0.01). Native T1 and late gadolinium enhancement radiomics demonstrated modest discrimination. Radiomics PCs showed associations with markers of cardiac size and remodeling, suggesting that radiomics captures morphologic and functional aspects (Figure 2). Correlations were modest, implying that radiomics are not simple surrogates of single conventional parameters but capture a broader spectrum of myocardial remodeling. Conclusion CMR-derived cine radiomics improved non-invasive detection of biopsy-proven MF in severe aortic stenosis, supporting its potential role in precision phenotyping and personalised management.Figure 1.Study workflowFigure 2.Correlation matrix