Towards joint segmentation and registration of the myocardium in CP-BOLD MRI at rest

İlkay Öksüz, Rohan Dharmakumar, Sotirios A. TsaftarisJournal of Cardiovascular Magnetic Resonance 2016 · DOI

Introduction

Cardiac Phase-resolved Blood Oxygen-Level Dependent (CP-BOLD) MRI is a promising, non-invasive technique for detecting myocardial ischemia without the need for exogenous contrast agents or physical stress. By monitoring subtle signal intensity changes throughout the cardiac cycle, clinicians can theoretically identify areas of reduced blood flow. However, the practical application of CP-BOLD is hindered by low image contrast and signal fluctuations, which make the essential tasks of myocardium segmentation and image registration extremely difficult. Standard tools often fail because they treat these two tasks as independent, sequential steps, leading to a "catch-22" where poor segmentation degrades registration and vice versa.

Methodology

To address these challenges, this research proposes a joint framework that treats segmentation and registration as a unified optimization problem. The core of the approach lies in an iterative loop that allows each process to inform and refine the other.

Feature Extraction and Dictionary Learning

Rather than relying on raw pixel intensities, which are unreliable in CP-BOLD, the system extracts $9 \times 9$ patches and combines them with Histogram of Oriented Gradients (HOG) and Gabor features. These features are then used to train two distinct dictionaries using the K-SVD algorithm: one for the background ($D_B$) and one for the myocardium ($D_M$).

The Iterative Loop

  1. Initial Segmentation: Pixels are classified based on their sparse representation within the learned dictionaries.
  2. Registration: Images are aligned using sparse coefficients as a similarity metric, providing robustness against intensity variations.
  3. Refinement: The registration results are used to update segmentation labels, which in turn are used to retrain the dictionaries for the next iteration.
Iterative framework for joint segmentation and registration using dictionary learning and label fusion.

Experimental Setup

The framework was validated using a dataset of 10 canine subjects imaged on a 1.5T Siemens Espree scanner. Data was collected under two conditions: a healthy baseline and severe Left Anterior Descending (LAD) coronary artery stenosis to simulate ischemic conditions. Each cardiac cycle provided approximately 25 frames for analysis.

Results and Discussion

The performance of the proposed joint approach was evaluated using the Dice overlap score. The results demonstrate that the iterative refinement process significantly improves accuracy compared to standard registration methods like DRAMMS and MIND, particularly in the challenging ischemic state.

Comparison of Dice overlap scores for the proposed approach versus standard registration methods.

By using sparse coefficients rather than raw pixels for registration, the algorithm effectively ignores the BOLD-related intensity fluctuations that typically confuse alignment tools. Furthermore, the subject-specific adaptation of dictionaries allows the model to handle individual anatomical variations more effectively than generic models.

Conclusion

This joint segmentation and registration framework offers a robust solution for the analysis of CP-BOLD MRI. While the current study is limited by a small canine sample size and high computational demands, it establishes a clear path toward reliable, contrast-free diagnosis of heart disease. Future work will focus on optimizing the algorithm for clinical processing speeds and validating the approach in human clinical trials.

Summary updated

BibTeX
@article{oksuz2016towards,
  author = {İlkay Öksüz and Rohan Dharmakumar and Sotirios A. Tsaftaris},
  title = {{Towards joint segmentation and registration of the myocardium in CP-BOLD MRI at rest}},
  journal = {Journal of Cardiovascular Magnetic Resonance},
  year = {2016},
  doi = {10.1186/1532-429x-18-s1-w34},
}
Figure