Abstract Body (Do not enter title and authors here): Introduction/Background: Genetic cardiomyopathy (GEN-CM) has overlapping features of inflammation and arrhythmias with cardiac sarcoidosis (CS) and non-genetic myocarditis (MYOC); however, the treatments are very different, making it critical to distinguish them accurately. Radiomics is an AI-based approach to pixel-based analysis for CMR images with the potential for broad implementation for this diagnostic problem using publicly available software.
Research Questions/Hypothesis: We tested the hypothesis that CMR radiomics would have an area under the curve (AUC) of ≧ 0.8 for diagnosing GEN-CM relative to CS and MYOC.
Methods/Approach: The study design was a secondary analysis of an observational cohort from an academic medical center. We evaluated contrast-enhanced CMR images from 145 patients in the University of Minnesota CMR Registry with either GEN-CM, CS, or MYOC. Radiomics was applied to SSFP cine end-diastolic images and late gadolinium enhancement (LGE) images. Principal component analysis (PCA) generated combinations of radiomics features ordered based on the percent of variance explained, and the most predictive PCA predictors for the two models (Model 1: GEN-CM v. CS; Model 2: GEN-CM v. MYOC) were chosen. Bootstrapping (resampling with replacement) was used to generate 100 different samples from the original dataset, and receiver operating characteristic (ROC) curves with 95% confidence bands were used to report the AUC and 95% CI for each model.
Results/Data: Among 145 patients (age 43.2 ± 16.0 years, 31% female) who had GEN-CM (n=50), CS (n=47) or MYOC (n=48), the most prominent features influencing the diagnosis were the maximum mean absolute deviation in pixel intensity, the maximum run length non-uniformity (normalized), and the large area emphasis feature, which were all higher in GEN-CM v. CS and GEN-CM v. MYOC (Figure 1). Four PCA predictors (three LGE predictors and one cine predictor) were identified for Model 1, and four PCA predictors (two LGE predictors and 2 cine predictors) were identified for Model 2. The AUC for Model 1 was 0.90 (95% CI 0.83 to 0.95), and the AUC for Model 2 was 0.88 (AUC 0.81 to 0.94), indicating a high level of performance (Figure 2).
Conclusions: Radiomics applied to CMR cine and LGE images provides a high level of performance for the diagnosis of genetic cardiomyopathy, accurately distinguishing it from cardiac sarcoidosis and non-genetic myocarditis with contributions from both CMR imaging sequences.