Feature Mapping of Native Oxygenation-Sensitive CMR Images for Classifying Cardiomyopathies

Cardiovascular disease remains a leading global health concern, necessitating innovative needle-free diagnostic tools. This study explores the integration of oxygenation-sensitive cardiovascular magnetic resonance (OS-CMR) imaging with deep learning to classify myocardial pathology across four categories: ischemic (42 cases), non-ischemic (33 cases), inflammation/edema (47 cases), and healthy myocardium (68 cases). Following image preprocessing and augmentation, the model was trained and evaluated using a stratified 5-fold cross-validation with Monte Carlo Dropout and residual learning. The final model achieved class-specific AUC scores of 0.93 (healthy), 0.80 (ischemic), 0.89 (non-ischemic), and 0.96 (edema) on the test dataset. Beyond classification, the layer activation maps were visualized and compared with expert-defined regions on LGE and T2 maps as interpretability tools. AI-derived feature maps demonstrated spatial correspondence with expert-defined lesions (Dice values: 0.85 for transmural ischemia, 0.90 for subendocardial involvement, 0.83 and 0.93 for non-ischemic lesions in HCM and DCM, and 0.93 for global edema). These findings suggest that the OS-CMR contains latent phenotype-specific information that can be leveraged by deep learning to support diagnostic classification. This may also allow for a comprehensive, ultra-efficient and needle-free CMR workflows in the future.