Novel Ultrasound Texture Based Similarity Metric Using Autoregressive Modelling.

Designing an ultrasound (US) specific similarity metric is essential in integrating advanced techniques like image segmentation and registration to US based interventional procedures. Applying conventional similarity metrics to ultrasound images is hampered by intrinsic noise patterns in an US image. In this work, we propose a texture based similarity metric (TexSimAR) using Autoregressive (AR) modelling. The key idea is to treat an US image as data resulting from a dynamical process which can be parametrically modelled. Using this approach it is possible to compute a parametric spectrum of individual US images and subsequently use it to estimate a similarity value between them. For evaluation, we used thyroid US images and similarity values were calculated between thyroid and non-thyroid regions. A cost function was designed to compare TexSimAR with other conventional similarity metrics. TexSimAR clearly distinguished between thyroid and non-thyroid regions outperfoming the conventional similarity metrics.