Utilizing Radiomics and Deep Learning to Predict Post-Contrast Information from Contrast Agent-Free Cine Cardiac Magnetic Resonance Imaging

Differentiating myocardial scar tissue from healthy myocardium and imaging artifacts is essential in clinical practice. This study investigates the feasibility of using radiomics and deep learning (DL) techniques to distinguish myocardial scar tissue from healthy myocardium and artifacts in contrast-free cine cardiac magnetic resonance (CMR) images. The dataset included 81 CMR images, divided into 65 for training and 16 for testing. A U-Net model was utilized to segment the myocardium and left ventricle (LV). U-Net segmentation performance was evaluated using the Dice Similarity Coefficient (DSC) and robust Hausdorff Distance (HD). Radiomics features extracted from the images trained machine learning classifiers to detect scar regions. During inference, these classifiers produced voxel-wise probability maps, aggregated for patient-level classification, identifying patients as either normal or having scars. In training cases, the mean DSC values were 0.87 for myocardium and 0.96 for LV cavity, with HD values of 1.03 and 1.02 pixels, respectively. For testing, DSC values were 0.84 for myocardium and 0.95 for LV cavity, with HD values of 1.15 and 1.0 pixels. The Gradient Boosting Classifier achieved an AUC ROC of 0.99–1.00 for training and 0.73 to 0.75 for testing. Patient-level classification AUC ROC ranged from 0.76–0.84 for training and 0.62–0.75 for testing. The classifier reached a sensitivity of 67% and specificity of 71% in testing. This automated approach holds promise for improving clinical workflows, aiding diagnosis, and enhancing prognosis by efficiently identifying myocardial scarring.