Introduction
Assessing tissue perfusion—the delivery of blood to an organ—is vital for diagnosing conditions such as coronary artery disease (CAD). Contrast-enhanced ultrasound (CEUS) offers a low-cost, non-invasive method for this assessment, yet interpreting the resulting image sequences remains a significant challenge. Clinical practice often relies on subjective visual interpretation, which is prone to inter-observer variability. Existing quantitative techniques frequently analyze pixels in isolation, ignoring spatial context, or apply ad hoc smoothing that can obscure fine temporal details. This paper presents a novel spatio-temporal framework that automatically classifies tissue perfusion by integrating Bayesian Factor Analysis (BFA) with Markov Random Fields (MRF), providing a robust and objective mapping of blood flow.
Methodology
The proposed approach treats tissue classification as a global optimization problem, simultaneously modeling temporal dynamics and spatial relationships.
Bayesian Factor Analysis (BFA)
BFA is employed to handle the temporal dimension of the data. It assumes that the observed intensity-time curve at each pixel is a linear combination of a small number of underlying "hidden factors," representing distinct perfusion characteristics (e.g., healthy tissue vs. blood pool). Unlike traditional Factor Analysis of Dynamic Studies (FADS), which often requires manual constraints to achieve a unique solution, this Bayesian formulation allows for an automated, objective estimation of the factor loading matrix and scores.
Markov Random Field (MRF)
To incorporate spatial context, the model utilizes a Markov Random Field. This component encodes the prior knowledge that neighboring pixels are likely to belong to the same tissue class, effectively regularizing the classification. The factor scores from the BFA serve as the posterior probabilities for the MRF. The entire system is solved using an Iterated Conditional Modes (ICM) algorithm, which iteratively refines tissue labels and model parameters until convergence.
Datasets
The framework was validated using:
- Simulated Data: Sequences based on an exponential replenishment model ($y_i = A(1 - e^{-\beta t})$) with three distinct flow rates ($\beta$) and varying levels of Gaussian noise (10% to 45%).
- In-Vivo Data: A myocardial perfusion study from a healthy volunteer, consisting of 14 ECG-triggered images to minimize motion artifacts.
Results and Discussion
The integration of spatial and temporal modeling yielded significant improvements in classification accuracy, particularly in high-noise environments typical of ultrasound imaging.
Noise Resilience and Accuracy
In simulations with 15% noise, the combined BFA-MRF model achieved a Misclassification Ratio (MCR) of just 1.94%. In contrast, using BFA alone resulted in an MCR of 8.99%, demonstrating that spatial regularization effectively mitigates "salt and pepper" noise. Even at an extreme noise level of 45%, the model maintained a reasonable MCR of 36.5%, successfully identifying primary tissue structures where pixel-wise methods would fail.
Clinical Plausibility
When applied to in-vivo cardiac data, the algorithm automatically segmented the image into three physiologically relevant regions: the myocardium (normal flow), the left ventricular cavity (blood pool), and areas of ultrasound attenuation (shadowing). This demonstrates the model's ability to reconstruct characteristic intensity-time curves that have clear physiological interpretations.
Conclusion
This spatio-temporal framework offers a promising path toward objective, automated perfusion mapping. By synergizing BFA and MRF, the method overcomes the limitations of subjective visual assessment and noise-sensitive pixel-wise analysis. Future work will focus on clinical validation in patients with coronary stenosis to evaluate the model's ability to detect and localize ischemic or infarcted regions compared to the current gold standard of manual semi-quantitative Region of Interest (ROI) analysis.
Summary updated