Late Gadolinium Enhanced Cardiac Magnetic Resonance Imaging Radiomics For High Precision Differentiation of Scar and Viable Cardiac Tissues

The aim of this work is to determine the radiomic features and develop a machine learning algorithm for high precision differentiation between scar and viable tissues in the left ventricular myocardium on Late Gadolinium Enhanced Cardiac Magnetic Resonance (LGE-CMR) images. This work involves patients referred for post-myocardial infarction (MI) with scar tissue in their left ventricles. Forty-two patients were included in the current study. All images were segmented and verified simultaneously by two radiologists using 3D-Slicer software. Radiomic features including intensity-based, texture-based and shape-based features were extracted from different regions. Hierarchical clustering was performed for unsupervised grouping of scar and viable tissues that differentiate them with one error in normal tissue. Multiple support vector machine recursive feature elimination (MSVM-RFE) was used for feature selection. For classification purposes, Support Vector Machine (SVM) and Random Forest (RF) were used in this study. Linear SVM with AUC of 0.99 ± 0.02, sensitivity of 0.99 ± 0.02 and specificity of 0.99 ± 0.02 yielded the best results. This work demonstrated that using radiomics on LGE-CMR images could accurately detect the scar tissue, which could potentially incorporate automated segmentation and classification tools in LGE-CMR.