PhD candidate McGill University Montréal, Quebec, Canada
Background: Accurate detection of myocardial scars and edema is essential for diagnosing various cardiomyopathies, including ischemic, non-ischemic, and inflammatory conditions. Cardiovascular magnetic resonance (CMR) is the gold standard for tissue characterization, with T1- and T2-weighted imaging and late gadolinium enhancement (LGE) providing complementary diagnostic insights. However, these sequences have limitations: LGE requires contrast administration and extended scan times, while T1 and T2 mapping are constrained by motion sensitivity, standardization issues, and susceptibility to artifacts. Oxygenation-sensitive CMR (OS-CMR) is a novel, contrast-free imaging technique based on breathing maneuvers, offering a non-invasive alternative. In this study, we introduce DeepOxyMap, an advanced AI-based framework that performs both classification and feature map visualization of OS-CMR images. For the first time, DeepOxyMap reveals scar and edema patterns directly from native OS-CMR, demonstrating that these images inherently encode pathological information typically attributed to T1 and T2 sequences. The aim of this work is to uncover these hidden diagnostic patterns to reduce dependence on contrast-enhanced and multi-sequence imaging.
METHODS AND RESULTS: This retrospective study analyzed 190 OS-CMR cases classified as ischemic scar (42), non-ischemic scar (33), myocardial edema (47), and healthy (68). Preprocessing included normalization, cropping, and augmentation. A VGG19-based convolutional neural network was adapted using transfer learning and trained to distinguish the four categories. Our DeepOxyMap model achieved 82% test accuracy, 86% precision, and 78% recall. ROC-AUC analysis showed excellent performance: 0.96 (healthy), 0.89 (ischemic), 0.95 (non-ischemic), and 0.97 (edema), with a micro-average of 0.96 (Figure1). The comparison with ResNet50 (87%) and EfficientNetB0 (79%) confirmed a higher performance of VGG19. Novel AI-generated heatmaps revealed that OS-CMR images contain tissue-specific features that closely align with ground-truth LGE, T1, and T2 maps, thereby validating that OS-CMR inherently encodes clinically relevant myocardial tissue pathology (Figure 1).
Conclusion: Our custom AI algorithm DeepOxyMap can classify patients according to their myocardial pathology based on native Oxygenation-Sensitive CMR images. The AI model also effectively localizes myocardial damage. If confirmed, this enables an ultra-efficient, needle-free CMR scan for LV function, tissue pathology, and coronary vascular function.
