Cardiomyopathies are a significant and heterogeneous group of diseases, with multiple classification schemes. Late Gadolinium Enhancement (LGE) Cardiovascular Magnetic Resonance (CMR) imaging is the standard for assessing focal myocardial fibrosis both of ischemic and non-ischemic origin. In order to guide disease specific device and medical therapy, clinicians examine the presence and extent of LGE myocardial damage with various mappings.
CMR images (n=2,701) representing 377 patients at The Royal Brompton Hospital Trust, where 219 patients (58%) have ischemic cardiomyopathy, were included in this study in order to develop and test two models to discriminate and identify ischemic cardiomyopathy from non ischemic dilated cardiomyopathy. The first model was human based contouring of infarct scar morphology, and a collection of biomarkers derived from it. The second model was built with a convolutional neural network to learn the scar features without any human interaction.
The first model found significant structural differences in scar formations between ischemic cardiomyopathy and non ischemic dilated cardiomyopathy, and achieved a reasonable classification performance AUROC 0.87 (± 0.02). The second model does not provide any further insights about the disease, but increases the discriminant power to AUROC 0.90 (± 0.05) in a cross-validation test. Future work is needed to achieve an optimal combination of both approaches in order to maximise the diagnostic capability without human variability and with interpretability.