DeepOxyMap: AI-Enhanced OS-CMR for Automated Detection of Local and Global Patterns of Myocardial Edema and Fibrosis

Background: Cardiovascular disease remains a leading global health concern, underscoring the need for accurate, efficient, and non-invasive diagnostic tools. Cardiac magnetic resonance imaging (CMR) is considered the gold standard for myocardial tissue characterization; however, each protocol requires time, patient cooperation, and often contrast injection, which may not be ideal for every patient [1]. Oxygenation-sensitive CMR (OS-CMR) can reveal vascular and metabolic abnormalities, particularly in heart failure patients, without the need for contrast agents [2,3]. Nevertheless, its interpretation for detailed tissue characterization remains challenging. We present DeepOxyMap, a deep learning framework that automatically classifies and visualizes both local and global myocardial pathology patterns, including ischemic scars, non-ischemic scars, and myocardial edema, from native OS-CMR images.

Methods: A VGG19-based convolutional neural network, enhanced with residual connections, Monte Carlo Dropout, was trained on preprocessed OS-CMR cine images from 190 subjects: ischemic scar (42), non-ischemic scar (33), edema/inflammation (47), and healthy myocardium (68). Data augmentation and stratified 5-fold cross-validation were applied. Beyond classification, pixel-level feature maps were extracted and binarized to localize pathology. These maps were compared to expert countering of abnormality on LGE, native T1, and T2 maps using Dice, IoU, Recall, and Precision.

Results: The model achieved an average classification accuracy of 81.7%, precision of 83.7%, recall of 78.9%, and AUC of 0.8552. ROC-AUC scores per class were 0.96 (healthy), 0.89 (ischemic), 0.95 (non-ischemic), and 0.97 (edema) (Figure 1). Feature map analysis demonstrated strong spatial correspondence with expert-defined regions: Dice scores of 0.8520 (large transmural ischemic scar), 0.8989 (subendocardial ischemic scar), 0.8333 (ICMP with HCM), 0.9299 (NICMP due to DCM phenotype), 0.8182 (localized small edema, Recall 0.999), and 0.9340 (persistent global myocardial edema) (Figure2,Table1). Second-reader validation showed strong contour agreement with Reader 1 (median Dice = 0.795, IQR 0.640–0.895). Even subtle pathologies, such as localized edema, were detected with near-perfect recall, demonstrating the model’s sensitivity to minimal myocardial involvement.

Conclusion: DeepOxyMap accurately classifies myocardial pathology and localizes both local and subtle patterns using native OS-CMR images. High similarity metrics across diverse pathology types confirm its ability to replicate expert-level segmentation without the need for contrast agents. This approach has the potential to improve diagnosis, particularly of edema f myocardial injury and fibrosis —without the need for contrast injection, while streamlining CMR workflows, reducing scan time, and enabling comprehensive, needle-free myocardial tissue characterization in a single acquisition.

Table 3. This analysis enabled the identification of the most interpretable and pathology-consistent feature maps.

Case Type	Dice	IoU	Recall	Precision
Large transmural Ischemic scar (IS)	0.8520	0.7421	0.8613	0.8429
Subendocardial Ischemic scar	0.8989	0.8163	0.9302	0.8696
Cardiomyopathy (ICMP)with HCM	0.8333	0.8333	0.8333	0.8333
NICMP group due to DCM phenotype	0.9299	0.8690	0.9012	0.9605
localized small edema	0.8182	0.6923	0.999	0.6923
Persistent global myocardial edema	0.9340	0.8762	0.9200	0.9485

Figure 1. Overview of the DeepOxyMap AI workflow for detecting and classifying myocardial tissue pathology in oxygenation-sensitive cine CMR (OS-CMR) images. (A) OS-CMR images are acquired from patients with various myocardial conditions and reviewed by experts for quality and diagnostic labeling. (B) The DeepOxyMap model uses the best-performing backbone (VGG-19) with additional residual blocks, Monte Carlo Dropout, and stratified K-Fold cross-validation. Model training and evaluation are illustrated with ROC-AUC plots and architecture diagrams. (C) Additional validation steps include generating AI heatmaps, binarizing them, and performing pixel-wise comparisons with expert-annotated ground truths from LGE, T1, and T2 maps to assess both classification accuracy and spatial localization. (D) AI-generated heatmaps are compared with binary expert masks, and similarity metrics are computed. The results demonstrate the model’s ability to localize pathological patterns for ischemic (ICMP), non-ischemic (NICMP), and inflammatory (edema) conditions, showing strong visual agreement with reference diagnostic maps.
Figure 2. compares similarity metrics across three different classes, A) Ischemic and non-ischemic B) edema patterns