The main task of computer-aided diagnosis (CADx) is to differentiate the pathological stages to which each detected colorectal lesion belongs, especially to differentiate hyperplastic polyps, which are non-neoplastic and seldom show malignant potential, from neoplastic lesions, which are malignant or at risk for malignant transformation. If we could extract useful pattern information from detected lesions, we would achieve the goal of the CADx task. In this paper, we aim to minimize the spatial variation in expanding the well-known Haralick texture descriptor in three-dimensional (3D) space for extraction and selection of volumetric texture features. Haralick et al described a way to compute measures along four directions in an image slice and select the mean and range over the four directions as the texture features. When extend their description in 3D space, we will have 13 directions and the feature selection would be the mean and range over the 13 directions. However, because of the heterogeneity of lesions' texture orientation, the mean and range over spatially variant directions may be sub-optimal. To mitigate the variation, we propose to perform one kind of principal component (PC) analyses, i.e., the Karhunen-Loeve transform, on the 13 directions and select the features along the PCs, instead of the mean and range.