Impact of Physical Noise Modeling on Image Segmentation in Echocardiography

Abstract

Segmentation is an essential task in ultrasound image analysis. Recently, the trend in literature is towards incorporation of high-level information, e.g., shape priors, since many low-level segmentation techniques suffer from the characteristics of medical ultrasound images, i.e., speckle noise, scattering artifacts, and shadowing effects. However, the majority of these works implicitly assume an additive Gaussian noise model in ultrasound images, although a strong deviation from this assumption is well known, and the impact of correct physical noise modeling is not examined sufficiently until now. In this paper we investigate the influence of three different noise models from literature using a variational region-based segmentation framework, which allows for the incorporation of both low-level and high-level information. We demonstrate that correct physical noise modeling is of high importance for the computation of accurate segmentation results. The numerical results are validated on real patient datasets from echocardiographic examinations and compared to manual segmentations from echocardiographic experts.