MR image segmentation of the knee bone using phase information

Magnetic resonance (MR) imaging is a widely available and well accepted non invasive imaging technique. Development of automatic and semi-automatic techniques to analyse MR images has been the focus of much research and numerous publications. However, most of this research only uses the magnitude of the acquired complex MR signal, discarding the phase information. In MR, the phase relates to the magnetic properties of tissues, information which is not found in the magnitude signal. As a result, phase is a complement to the magnitude signal and can improve the segmentation and analysis of MR images. In this paper, we consider the automatic classification of textured tissues in 3D MRI. Specifically, we include features extracted from the phase of the MR signal to improve texture discrimination in the bone segmentation. Our approach does not require phase unwrapping, with the MR signal processed in its complex form. The extra information extracted from the phase provides better segmentation, compared to only using magnitude features. The segmentation approach is integrated within a novel multiscale scheme, designed to improve the speed of pixel based classification algorithms, such as support vector machines. An order of magnitude increase is obtained, by reducing the number of pixels that need to be classified.