Artificial-intelligence-based feature mapping of native oxygenation-sensitive cardiovascular magnetic resonance...

Cardiovascular magnetic resonance (CMR) is the non-invasive gold standard for myocardial tissue characterization. This, however, requires additional sequences and typically the use of contrast agents. Native oxygenation-sensitive cardiovascular magnetic resonance (OS-CMR) has been introduced for assessing coronary vascular function. The resulting images may allow for the extraction of tissue information. We hypothesized that deep learning can extract phenotype-specific information from OS-CMR images to enable contrast-free classification of myocardial pathology.

We developed and trained a deep learning algorithm to extract tissue information for classifying myocardial pathology.

We developed and trained a deep learning algorithm to extract tissue information for classifying myocardial pathology. We used data from 190 individuals classified as follows: Ischemic (n=42), non-ischemic (n=33), inflammation/edema (n=47), and healthy (n=68). We divided the scans into sets for training (70%), validation (20%), and testing (10%). The model was evaluated using a stratified five-fold cross-validation with Monte Carlo Dropout and residual learning. Beyond classification, the regions and extent of the layer activation maps were compared with results from blinded expert reads. Spatial agreement was quantified using Dice similarity coefficients.

The model achieved class-specific area under the curve (AUC) values of 0.93 for healthy myocardium, 0.80 for ischemic cardiomyopathy, 0.89 for non-ischemic cardiomyopathy, and 0.96 for inflammation/edema. Artificial intelligence-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 hypertrophic cardiomyopathy and dilated cardiomyopathy, and 0.93 for global edema).

OS-CMR images contain phenotype-specific information that can be extracted by deep learning to support a diagnostic classification of myocardial tissue pathology. The multi-parametric analysis may enable a comprehensive, ultra-efficient and needle-free CMR scan.