Extracting tissue pathology from native oxygenation-sensitive CMR images using artificial intelligence

Background Accurate detection of myocardial scars and edema is essential for diagnosing ischemic, non-ischemic, and inflammatory injury. While CMR with T1, T2, and LGE provides detailed tissue characterization, these methods are limited by contrast use, motion sensitivity, and long scan times. OS-CMR offers a contrast-free, breathing-based alternative. We present DeepOxyMap, a novel AI framework that classifies and extracts features maps from native OS-CMR images, revealing diagnostic patterns without contrast agents. Purpose We aimed to investigate the ability of a novel deep learning model to identify myocardial tissue pathology in native oxygenation-sensitive CMR images. Methods We retrospectively analyzed 190 OS-CMR cases (42 ischemic, 33 non-ischemic, 47 edema, 68 healthy) from two cohorts. Images were preprocessed and augmented using random rotations and horizontal flips. We tested several models: VGG19, ResNet50, EfficientNetB0, and VGG19 with Monte Carlo dropout and K-Fold cross-validation (MC+KF). The best-performing model, Residual VGG19 + MC+KF, was selected for further evaluation. Pixel-wise activation maps from layers were extracted by DeepOxyMap, binarized, and filtered to create heatmaps, which were compared with expert annotations from LGE (scars) and T2 maps (edema). Results Baseline VGG19 achieved 80.4% accuracy, while ResNet50 and EfficientNetB0 performed around 60%. VGG19+MC+KF showed improved stability across five folds (81.0% accuracy, 87.2% precision, 77.2% recall). Residual VGG19+MC+KF outperformed all, reaching 81.7% accuracy, 83.7% precision, 79.0% recall, and 0.86 AUC (Figure 1). To enhance model interpretability, we extracted pixel-level feature maps from convolutional layers and directly evaluated their spatial correspondence with expert-defined pathological regions. Unlike Grad-CAM, this approach produced pixel-level, high-resolution maps highlighting pathological patterns. These AI-maps were then compared against expert contours: LGE for scars, T2 for edema. Pixel-wise evaluation demonstrated high spatial overlap, with Dice scores of 0.90 (non-ischemic), 0.84 (ischemic), and 0.71 (edema). Pearson correlation coefficients reached up to 0.88, confirming robust alignment between learned features and ground-truth pathology (Figure 2). Conclusion DeepOxyMap demonstrates the ability to extract diagnostic markers of myocardial pathology from native OS-CMR images with high classification accuracy and spatial precision. This method allows for non-invasive detection of ischemic, non-ischemic, and inflammatory conditions without the need for contrast agents or additional sequences, potentially streamlining clinical workflows and reducing scan time and cost.