Segmentation of the Bones in MRIs of the Knee Using Phase, Magnitude, and Shape Information

Abstract

The segmentation of textured anatomy from magnetic resonance images (MRI) is a difficult problem. We present an approach that uses features extracted from the magnitude and phase of the MRI signal to segment the bones in the knee. Moreover, we show that by incorporating shape information, more accurate and anatomically valid segmentations are obtained. Eighteen volunteers were scanned in a whole-body 3T clinical scanner using a transmit-receive quadrature extremity coil. A gradient-echo sequence was used to acquire three-dimensional (3D) volumes of raw complex image data consisting of phase and magnitude information. These images were manually segmented and features were extracted using a bank of Gabor filters. The extracted features were then used to train a support vector machine (SVM) classifier. Each image was then automatically segmented using both the SVM classifier and a 3D active shape model (ASM) driven by the classifier. The use of phase and magnitude information from both echoes obtained the most accurate classifier results with an average dice similarity coefficient of 0.907. The use of 3D ASMs further improved the robustness, accuracy and anatomic validity of the segmentations with an overall DSC of 0.922 and an average point to surface error along the bone-cartilage interface of 0.73 mm. Our results demonstrate that the incorporation of phase and multiple echoes improve the results obtained by the classifier. Moreover, we show that 3D ASMs provide a robust and accurate way of using the classifier to obtain anatomically valid segmentation results.