Quantification of Myocardial Perfusion Lesions using Spatially Variant Finite Mixture Modelling of DCE-MRI

Yalei Yang, Hao Gao, Colin Berry, Aleksandra Radjenovic, Dirk HusmeierProceedings of the International Conference on Statistics, Theory and Applications (ICSTA ...) 2019 · DOI

Abstract

Dynamic Contract Enhanced Magnetic Resonance (MR) Imaging (DCE-MRI) can reveal differences in myocardial perfusion (microvascular or capillary blood flow) within the myocardium. The detection and quantification of hypo-perfused lesions within the myocardium is important for understanding the aetiology of coronary heart disease (CHD). In this paper, a modification of a traditional method, the Expectation-Maximization (EM) algorithm for Gaussian Mixture Models (GMM), is implemented. This modification, the Spatially Variant Finite Mixture Model (SVFMM), is able to take the neighbourhood information of a voxel in the MR image into account. An experiment based on both synthetic and real images illustrates and quantifies the improvement achieved with SVFMM over the traditional GMM method.

Introduction

Coronary Heart Disease (CHD) often leads to myocardial ischemia—a condition where blood flow to the heart muscle is restricted. Clinicians utilize Dynamic Contrast-Enhanced MRI (DCE-MRI) to visualize these "hypo-perfused" lesions. However, accurately quantifying these areas is technically challenging due to the low Signal-to-Noise Ratio (SNR) often found in rapid cardiac imaging sequences.

The Spatial Context Problem

Standard statistical approaches, such as the Gaussian Mixture Model (GMM), classify pixels based solely on their signal intensity. The primary drawback of this method in MRI analysis is the production of "island" pixels—single pixels labeled as diseased while surrounded by healthy tissue.

Physiologically, myocardial lesions are typically large and contiguous because they follow the distribution of major coronary arteries; scattered "island" pixels are usually imaging artifacts. To improve accuracy, the classification algorithm must incorporate the prior knowledge that a pixel is more likely to be part of a lesion if its neighbors are also part of a lesion.

Spatially Variant Finite Mixture Modelling (SVFMM)

The researchers implemented a Spatially Variant Finite Mixture Model (SVFMM) to address these limitations. Unlike the standard GMM, SVFMM introduces a Gibbs function as a spatial prior. This acts as a "smoothness" constraint, penalizing the model when adjacent pixels are assigned to different groups unless the signal intensity evidence is exceptionally strong.

By using Maximum A Posteriori (MAP) estimation, the model balances data fit with this spatial prior. A regularization parameter, $\beta$, was introduced to control this balance: a higher $\beta$ places more importance on neighborhood information, effectively smoothing the resulting parametric map.

Experimental Validation

The study utilized two primary data sources to validate the model:

  1. Clinical Data: Derived from a clinical cohort study involving patients with ST-segment elevation myocardial infarction (STEMI), focusing on short-axis slices captured during the "first pass" of a contrast agent.
  2. Synthetic Data: A synthetic myocardium ring was created to provide a "ground truth." It was split into healthy and diseased halves with added noise to mimic the exact SNR found in clinical images.

Results and Discussion

The SVFMM significantly outperformed the traditional GMM in both datasets by reducing physiologically improbable artifacts and improving boundary definition.

Visual comparison of segmentation results for clinical (left) and synthetic (right) data. SVFMM with higher beta values produces more contiguous regions compared to the noisy GMM output.

In clinical data, where the true lesion size is unknown, the authors used the count of isolated "island" pixels as a proxy for error. As shown in the table below, the standard GMM resulted in 59 isolated pixels. By applying SVFMM with an optimal $\beta=20$, that number dropped to 25—a more than 50% reduction in likely artifacts.

Quantitative comparison of isolated 'island' pixels between SVFMM and standard GMM.

In the controlled synthetic test, the standard GMM misclassified 55 pixels, while the SVFMM (at $\beta=20$) misclassified only 10 pixels. The resulting parametric maps showed much cleaner, more distinct boundaries between healthy and diseased tissue, enhancing their utility for clinical decision-making.

Conclusion and Future Work

By incorporating spatial context, the SVFMM provides a more robust method for quantifying myocardial lesions than standard intensity-based models. Future research will focus on automating the selection of the regularization parameter $\beta$ through Bayesian evidence optimization and refining the model to better handle "signal bleeding" at the boundaries of the heart wall.

Summary updated

Abstract

Dynamic Contract Enhanced Magnetic Resonance (MR) Imaging (DCE-MRI) can reveal differences in myocardial perfusion (microvascular or capillary blood flow) within the myocardium. The detection and quantification of hypo-perfused lesions within the myocardium is important for understanding aetiology of coronary heart disease (CHD). In this paper, a modification of a traditional method, the Expectation-Maximization (EM) algorithm for Gaussian Mixture Models (GMM), is implemented. This modification, the Spatially Variant Finite Mixture Model (SVFMM), is able to take the neighbourhood information of a voxel in the MR image into account. An experiment based on both synthetic and real images illustrates and quantifies the improvement achieved with SVFMM over the traditional GMM method.
BibTeX
@inproceedings{yang2019quantification,
  author = {Yalei Yang and Hao Gao and Colin Berry and Aleksandra Radjenovic and Dirk Husmeier},
  title = {{Quantification of Myocardial Perfusion Lesions using Spatially Variant Finite Mixture Modelling of DCE-MRI}},
  booktitle = {Proceedings of the International Conference on Statistics, Theory and Applications (ICSTA ...)},
  year = {2019},
  doi = {10.11159/icsta19.26},
}
Figure