Towards pixel-wise area-at-risk characterization with cardiac BOLD MRI at rest

Marco Bevilacqua, Rohan Dharmakumar, Sotirios A. TsaftarisJournal of Cardiovascular Magnetic Resonance 2016 · DOI

Introduction

Identifying the "area-at-risk"—the myocardial region vulnerable to infarction due to restricted blood flow—is a cornerstone of cardiac diagnostics. Conventional MRI techniques often necessitate the use of exogenous contrast agents (e.g., Gadolinium) or pharmacological stress tests to simulate exercise. While Cardiac Phase-resolved Blood-Oxygen-Level-Dependent (CP-BOLD) MRI offers a non-invasive alternative by monitoring oxygenation changes throughout the cardiac cycle at rest, current analysis methods are often limited to coarse segmental averaging. This research introduces a machine learning framework designed to move toward pixel-wise characterization by integrating both signal intensity and mechanical wall motion.

Methodology

The proposed approach utilizes a pattern recognition framework that treats cardiac behavior as a time series. By analyzing CP-BOLD MRI data, the researchers focused on two synergistic data streams:

  1. Signal Intensity: A proxy for local blood oxygenation levels.
  2. Wall Motion: Quantified as the average radial distance between the endocardial and epicardial boundaries across the cardiac cycle.

The myocardium is divided into 36 radial segments. An iterative algorithm employing Dictionary Learning and a One-Class Support Vector Machine (OC-SVM) is then applied. The model learns the manifold of "normal" myocardial behavior from remote regions; the OC-SVM subsequently flags segments deviating from this norm as ischemic. This process culminates in the generation of an "ischemia likelihood map."

Ischemia likelihood map (left) and bullseye plot (middle) generated by the proposed algorithm, compared against ground truth imaging (right).

Key Results

The study, conducted on a canine model ($n=8$) with induced LAD stenosis, compared the proposed method against traditional metrics like the Systole/Diastole (S/D) ratio and Independent Component Analysis (ICA). Late Gadolinium Enhancement (LGE) images served as the ground truth for infarct size.

  • Synergistic Performance: The integration of both intensity and wall motion data yielded a high Pearson correlation of $r = 0.84$ ($p = 0.004$) with the actual infarct size.
  • Intensity-Only Analysis: Even when restricted to signal intensity, the machine learning approach ($r = 0.66$) significantly outperformed traditional methods.
  • Limitations of Standard Techniques: Conventional approaches failed to provide significant correlations in the rest-state data, with the S/D ratio achieving only $r = -0.08$ and ICA reaching $r = 0.51$.

Discussion and Future Work

The primary breakthrough of this work is the demonstration that combining oxygenation (intensity) and mechanics (wall motion) provides a more comprehensive diagnostic signature than either metric alone. By using unsupervised outlier detection (OC-SVM), the system can identify pathology without requiring massive labeled datasets of diseased states.

However, several hurdles remain before clinical translation. The small sample size ($n=8$) and the use of a controlled canine model mean the robustness of the algorithm in human patients—who often present with complex multi-vessel disease and comorbidities—remains to be established. Additionally, the computational demands of iterative dictionary learning must be evaluated for real-time clinical workflows.

Summary updated

BibTeX
@article{bevilacqua2016towards,
  author = {Marco Bevilacqua and Rohan Dharmakumar and Sotirios A. Tsaftaris},
  title = {{Towards pixel-wise area-at-risk characterization with cardiac BOLD MRI at rest}},
  journal = {Journal of Cardiovascular Magnetic Resonance},
  year = {2016},
  doi = {10.1186/1532-429x-18-s1-p63},
}

Loading related content...

Figure