Quick Fire Session
SCMR 22nd Annual Scientific Sessions
Background: Late Gadolinium Enhancement (LGE) imaging is a reference standard technique for the differentiation of Ischemic Cardiomyopathy (ICM) from Non-Ischemic Dilated Cardiomyopathy (NIDCM) in patients with heart failure and reduced ejection fraction (HFrEF). However, advancements in 3D Myocardial Deformation Analysis (3D-MDA) offer reproducible multi-component phenotypic evaluations that may provide value for artificial intelligence (AI)-assisted cardiomyopathy diagnosis. In this study, we aimed to evaluate the role of Machine Learning (ML) to distinguish ICM from NIDCM using multiple geometrical (wall thickness) and functional (strain amplitude, time to systolic peak strain, strain rate) parameters derived from feature-tracking based 3D-MDA.
Routine, 2D-multiplanar cine and LGE imaging was acquired in 164 patients with a Left Ventricular Ejection Fraction (LVEF) <40%. All patients were classified using clinical and LGE-based criteria as having ICM (N=78) or NIDCM (N=86). LGE-based classifications were performed by blinded expert reader with coding of fibrosis presence and spatial distribution. 3D-MDA was performed using in-house validated software (GIUSEPPE) to obtain segmental and global values of strain amplitude, time to systolic peak and peak systolic rate in conventional (radial, circumferential, and longitudinal) and principal (maximum, minimum, and secondary) directions as well as spatial distributions of mean diastolic wall thickness. Regional and global features with the strongest discriminatory power were selected after performing a Principal Component Analysis (PCA). The principal components were used to train a logistic regression algorithm to discriminate ICM from NIDCM. In order to evaluate the quality of the learning system, we trained a model on 80% of the population, then tested it on the remaining 20%.
Mean age was 59.8±10.7 years with 45 female subjects (27%) and a mean LVEF of 26.7±7.6%. The respective LVEF in ICM and NIDCM patients were 26.9±6.8% and 26.5±8.3% (p=0.92). No significant differences were found between ICM and NIDCM groups for global strain amplitude (all directions), while global time to peak strain was different in all directions (p<0.05). Individually, global measures performed modestly for the classification of ICM versus NIDCM; the highest AUC was found for circumferential strain (AUC: 0.67) (Figure 1). A PCA-based logistic regression model using regional 3D-MDA achieved an AUC of 0.87 (sensitivity 82%, specificity 80%, PPV 82%, NPV 80%).
Conclusion: Machine learning based analysis of regional 3D-MDA data allows for robust discrimination of ICM versus NIDCM. The capacity of this technique to advance AI-assisted diagnosis warrants expanded investigation in larger, multi-centre cohorts.