Segmentation of atherosclerotic plaque components in ultrasonic B-mode images using a multiphase Bayesian level-set

Partitioning an atheromatous carotid plaque into its main biological components can be a valuable tool to assess its vulnerability. One could compute textural or elasticity features on each component to describe them. In this paper, we propose a fully automated segmentation method, based on statistical properties of the ultrasound backscattered envelope signals, to classify plaque pixels into a fixed number of components. The echogenicity of each plaque was modeled as a mixture of 2 Nakagami distributions leading to 2 main components. The mixture parameters were at first estimated with an Expectation Maximization algorithm (EM). Each class of the partition was then initialized using the Maximum Likelihood segmentation (ML). The optimal partition of the plaque area was then found using a level-set formulation of the Maximum A Posteriori (MAP) estimator with a spatial cohesion prior. Nakagami mixture parameters were extracted from those components. Uncompressed B-mode sequences of 8 symptomatic and 13 asymptomatic subjects were analyzed for a total of 42 plaque sequences. We found that the Nakagami parameters were able to distinguish symptomatic from asymptomatic patients with a significant p-value. Further works including elasticity mapping on each component are in progress and might lead to new indexes of vulnerability.