Abstract
Cardiac phase-resolved Blood-Oxygen-Level-Dependent (CP-BOLD) MRI can detect myocardial ischemia at rest without contrast and stress agents. At rest, BOLD myocardial signal intensity varies with cardiac phase: in healthy conditions it is maximal in systole and minimal in diastole, but in disease this pattern is not evident. These changes are not readily visible and post-processing is necessary. Previous methods used segmental analysis and only two images to identify ischemic segments. In this study we demonstrate that it is possible to use unsupervised learning methods to identify ischemia with higher accuracy.
Introduction
Detecting myocardial ischemia—a condition where heart muscle receives insufficient oxygen—traditionally requires invasive contrast agents or pharmacological stress tests. Cardiac phase-resolved Blood-Oxygen-Level-Dependent (CP-BOLD) MRI has emerged as a promising non-invasive alternative that monitors oxygen levels naturally throughout the cardiac cycle at rest.
In a healthy myocardium, the BOLD signal follows a predictable cyclic pattern, peaking during systole (contraction) and dipping during diastole (relaxation). Ischemia disrupts this pattern; however, these changes are often too subtle for visual detection. Current post-processing techniques, such as the Systolic to Diastolic Ratio (SDR), typically rely on only two snapshots in time, discarding the majority of the temporal data collected and resulting in limited sensitivity. This study proposes an unsupervised dictionary learning approach to leverage the full cardiac cycle for improved diagnostic accuracy.
Methods
The researchers developed a machine learning pipeline that treats the heartbeat as a complex signal processing problem. The myocardium is first segmented into 24 regions, and signal intensity for each is extracted across all phases of the cardiac cycle to create a 1D time series.
Instead of manual feature extraction, the team employed unsupervised dictionary learning. They modeled the signal ($Y$) as a sparse combination ($X$) of a dictionary ($D$) of patterns ($Y \approx DX$). A critical component of this dictionary is its circulant structure, which accounts for the cyclic phase shifts inherent in healthy heartbeats. The algorithm uses an iterative process involving a 1-class Support Vector Machine (SVM) to identify "normal" segments. The dictionary is then updated exclusively using these normal segments to ensure it accurately represents healthy tissue. The final output is an Ischemia Likelihood Value (ILV) between 0 and 1, providing a spatial map of potential ischemic territories.
Results
The proposed dictionary learning approach was validated using both synthetic datasets and canine models of severe Left Anterior Descending (LAD) artery stenosis. On synthetic data, the dictionary learning method achieved an accuracy of 92–94%, significantly outperforming the 72–74% accuracy of the traditional SDR method.
In canine models, the ILV maps correctly identified territories affected by stenosis. These findings were visually and spatially consistent with Late Gadolinium Enhancement (LGE) imaging, the current clinical gold standard for identifying myocardial damage. The results demonstrate that analyzing the entire cardiac cycle captures temporal nuances that two-point snapshots miss.
Discussion and Conclusion
By utilizing an unsupervised approach, this method learns healthy patterns automatically from the data, eliminating the need for large pre-labeled training sets and making it adaptable to various patient profiles and scanner settings. The inclusion of circulant dictionaries provides necessary shift-invariance to handle natural variations in cardiac phase timing.
While the results are promising, the authors acknowledge that further validation in large-scale human clinical trials is required. Future work will also focus on optimizing the computational complexity of the dictionary learning process to allow for real-time clinical integration. This technology holds potential for triaging patients with NSTEMI (non-ST-elevation myocardial infarction) in emergency settings where rapid, non-invasive diagnosis is critical.
Summary updated