A Probabilistic Framework for Ultrasound Image Decomposition

Image segmentation and tissue characterization are fundamental tasks of computer-aided diagnosis (CAD) in medical ultrasound imaging. As an initial step, such algorithms are usually based on extraction of pertinent features from the acquired ultrasound data. Typically, these features are computed directly from localized image segments, thereby representing local statistical properties of the image. However, the process of image formation of medical ultrasound suggests that such an approach could result in a variety of unwanted artifacts (such as excessively smooth segmentation boundaries or misclassification) at subsequent stages of the algorithm. In this work, we propose to first decompose the observed images into a number of their statistically distinct components. The decomposition is based on the maximum-a-posteriori (MAP) statistical framework which has been derived based on the signal and noise models appropriate for the ultrasound setting. Subsequently, each resulting component is used separately to extract a set of its corresponding features. When retrieved in this way (rather than directly from the observed image), the combined set of resulting features is shown to be capable of better discriminating between different tissue types. Examples of in silico simulations and in vivo experiments are provided to illustrate the practical usefulness of this technique for improving the results of ultrasound image segmentation