Improving feature extraction methods for ct texture analysis

In this paper we discuss the problem of discriminating tis sues with similar average Hounsfield values in Computed Tomography (CT) images through the use of supervised classification of feature vectors computed on small tex ture patches. We point out the differences between this problem and classical texture classification workbenches and analyze the role of data pre-processing (depth sub sampling, equalization) in determining how well classical texture features based on Gray Level Run Length Matri ces (GLRLM) and Gray Level Co-Occurrence Matrices (GLCM), discriminate tissues. Depth reduction and con trast stretching are shown to be key factors determining the information captured by features and can be interpreted as a “material segmentation”. Theory and experimental results show that different pre-processing does change error rates of supervised classifiers trained with GLRLM and GLCM based descriptors and that, using optimal depth subsam pling methods it is possible to obtain good texture classi fication results compatible with a physical interpretation of texture elements. On the basis of this interpretation, a new simple descriptor using ad hoc image thresholding and shape analysis is also introduced and compared with the previously discussed methods as well as with wavelet based multi-resolution filterin