Radiomics Score Derived from Parametric Mapping Cardiovascular Magnetic Resonance May Predict Left Ventricular Reverse Remodeling in Dilated Cardiomyopathy

Background: Achieving left ventricular reverse remodeling (LVRR) is a major therapeutic goal of guideline-directed medical therapy in patients with dilated cardiomyopathy (DCM). Assessment of myocardial tissue characteristics is essential for elucidating the pathophysiological mechanisms of DCM that are potentially linked to patient outcomes. Conventional cardiovascular magnetic resonance (CMR) techniques, including native T1 mapping and extracellular volume (ECV) mapping, provide valuable information but are limited in their ability to capture subtle and heterogeneous myocardial changes. Radiomics, by extracting high-dimensional features from medical images, offers a novel quantitative approach for a more comprehensive characterization of myocardial tissue [1]. We investigated whether radiomics features extracted from parametric mapping images on CMR can predict LVRR in DCM.

Methods: We retrospectively analyzed data from 75 DCM patients who underwent CMR at Osaka University Hospital from July 2017 to October 2024. LVRR was defined as left ventricular ejection fraction of ≥ 40% with ≥ 10% increase at least 180 days after CMR. Radiomics features were extracted from mid-septum on native T1 or ECV maps. The radiomics score (RadS) was computed for each case via a linear combination of selected radiomics features weighted by their respective coefficients calculated by LASSO‑penalized logistic regression with cross-validation. Univariate and multivariate logistic regression analyses of CMR variables and clinical data were performed to predict LVRR. We then developed prediction models for LVRR by incorporating independent predictors identified in the multivariate analysis. Receiver operating characteristic (ROC) curves were generated to evaluate the performance of each model.

Results: Patients with LVRR (n=31, 42%) had higher T1-based RadS (0.72 ± 1.52 vs. -1.54 ± 2.66, p< 0.0001), higher ECV-based RadS (0.24 ± 0.65 vs. -0.16 ± 0.64, p< 0.01), lower mean ECV (31.5 ± 3.8% vs. 34.9 ± 6.0%, p< 0.005), and lower extent of late gadolinium enhancement (%LGE) (0.2 ± 0.6% vs. 3.2 ± 6.5%, p< 0.05) than those without LVRR. Multivariate logistic regression analysis showed that only T1-based RadS (OR, 1.86; 95 % CI, 1.25–3.22; p=0.011) and ECV-based RadS (OR, 4.38; 95 % CI, 1.59–15.2; p=0.009) were independent predictors of LVRR. The area under the ROC curve (AUC) for predicting LVRR was 0.835 (95% CI, 0.712–0.910) for T1-based RadS and 0.718 (95% CI, 0.574–0.828) for ECV-based RadS. Adding the combination of T1-based and ECV-based RadS to the %LGE model significantly improved the AUC from 0.618 to 0.908 (p< 0.0001) (Figure, Table).

Conclusion: Radiomics scores extracted from native T1 and ECV maps may improve the prediction of LVRR and provide added value beyond traditional LGE assessment in patients with DCM.

AUCs of ROC Curves for Predicting LVRR.

Variables	AUC	95% CI	Sensitivity, %	Specificity, %	p
%LGE	0.618	0.48–0.74	58.1	58.5	reference
T1 RadS	0.834	0.72–0.91	67.7	75.6	0.007
ECV RadS	0.718	0.57–0.83	50.0	76.5	0.278
T1 RadS + ECV RadS	0.867	0.75–0.94	80.0	76.5	0.004
%LGE + T1 RadS	0.864	0.76–0.93	80.6	73.2	0.0002
%LGE + ECV RadS	0.800	0.67–0.89	63.3	76.5	0.009
%LGE + T1 RadS + ECV RadS	0.903	0.79–0.96	86.7	79.4	< 0.0001

Abbreviations: LVRR, left ventricular reverse remodeling; LGE, late gadolinium enhancement; ECV, extracellular volume;

RadS, radiomics score; AUC, area under the ROC curve; ROC, receiver operating characteristic; CI, confidence interval.ROC curves for predicting LVRR using CMR-derived variables (%LGE, T1 RadS, ECV RadS, and their combinations).