Abstract
Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) is a critical non-invasive tool for assessing myocardial blood flow (MBF). Identifying hypo-perfused regions—areas with restricted blood flow—is essential for diagnosing coronary heart disease. This paper introduces a Markov Random Field constrained Gaussian Mixture Model (GMM-MRF) to classify MBF maps. By incorporating spatial neighborhood information and utilizing the Iterated Conditional Modes (ICM) algorithm, the proposed method achieves high classification accuracy with significantly reduced computation time compared to traditional hierarchical models.
Introduction
In clinical cardiology, accurately delineating lesions in heart tissue is vital. While Gaussian Mixture Models (GMM) are standard statistical tools for clustering, they treat pixels as independent entities. In the context of noisy medical imaging, this often results in "speckled" maps containing physiologically impossible isolated clusters. While Hierarchical Bayesian Models (HBM) address this by considering spatial context, their computational cost—often requiring hours per scan—limits clinical utility. This research proposes the GMM-MRF as a computationally efficient "middle ground" that maintains spatial intelligence.
Methodology
The GMM-MRF model assumes that the myocardial tissue consists of two primary distributions: healthy and hypo-perfused (lesion). To enforce physiological realism, a Markov Random Field (MRF) prior is integrated. This constraint posits that a pixel's classification is influenced by its neighbors, favoring spatial continuity.
The model considers both first-order and second-order neighbors (as seen in the figure above). To optimize the model quickly, the researchers employed the Iterated Conditional Modes (ICM) algorithm. ICM is a greedy optimizer that iteratively updates labels to find a local maximum of probability, enabling processing times of less than a minute.
Experimental Results
The model was validated using both synthetic datasets (mimicking Gaussian and Rician noise) and clinical DCE-MRI scans.
Key findings include:
- Accuracy: In synthetic tests, GMM-MRF ($T=1$) reduced misclassified pixels from 64.0 (standard GMM) to just 3.8.
- Speed: The GMM-MRF method completed calibration and execution in under one minute, whereas the HBM benchmark required several hours.
- Spatial Weighting ($T$): The parameter $T$ controls the strength of the spatial prior. As shown in the figure above, lower values of $T$ (stronger priors) lead to smoother regions, while higher values allow for more noise.
Conclusion
The GMM-MRF provides a robust and rapid solution for myocardial perfusion classification. While the ICM algorithm's greedy nature can occasionally lead to local optima in the presence of severe imaging artifacts, the trade-off in speed makes it highly suitable for clinical environments. Future work will focus on incorporating spatial correlation directly into the noise component to further improve performance against MRI-specific artifacts.
Summary updated