Myocardial Perfusion Classification Using A Markov Random Field Constrained Gaussian Mixture Model

Yalei Yang, Hao Gao, Colin Berry, Aleksandra Radjenovic, Dirk HusmeierProceedings of the International Conference on Statistics, Theory and Applications (ICSTA ...) 2022 · DOI

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.

Neighborhood structure showing first-order (immediate) and second-order (diagonal) neighbors used to determine spatial consistency.

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.

Comparison of classification performance on synthetic data. Note the 'speckled' noise in the standard GMM (c) compared to the spatially consistent results of GMM-MRF (d-g) and the ground truth (b).

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

Abstract

Dynamic Contract Enhanced Magnetic Resonance (MR) Imaging (DCE-MRI) has been widely used as a non-invasive assessment approach to estimate the myocardial blood flow (MBF). The delineation of a hypo-perfused region (low MBF region) is important for understanding a patient's heart condition in clinical diagnosis. In this paper, a Markov random field constrained Gaussian mixture model (GMM-MRF) classification method is introduced to classify MBF maps using myocardial perfusion DCE-MRI data. The GMM-MRF method, trained with an ICM algorithm, makes use of spatial neighbourhood information to improve classification accuracy. The proposed method is applied to and assessed on both synthetic and clinical data, and compared with established classification methods.
BibTeX
@inproceedings{yang2022myocardial,
  author = {Yalei Yang and Hao Gao and Colin Berry and Aleksandra Radjenovic and Dirk Husmeier},
  title = {{Myocardial Perfusion Classification Using A Markov Random Field Constrained Gaussian Mixture Model}},
  booktitle = {Proceedings of the International Conference on Statistics, Theory and Applications (ICSTA ...)},
  year = {2022},
  doi = {10.11159/icsta22.146},
}
Figure