Extrapolation techniques for textural characterization of tissue in medical images

The low in-plane resolution of thoracic computed tomography (CT) scans may force texture analysis in regions of interest (ROIs) that are not completely filled by the tissue under analysis. The inclusion of extraneous tissue textures within the ROI may substantially contaminate these texture descriptor values. The goal of this study is to investigate the accuracy of different image extrapolation methods when calculating common texture descriptor values. Three extrapolation methods (mean fill, tiled fill, and CLEAN deconvolution) were applied to 480 lung parenchyma regions of interest (ROls) extracted from transverse thoracic CT sections. The ROls were artificially corrupted, and each extrapolation method was independently applied to create extrapolation-corrected ROIs. Texture descriptor values were calculated and compared for the original, corrupted, and extrapolation-corrected ROIs. For 51 of 53 texture descriptors, the values calculated from extrapolation-corrected ROls were more accurate than values calculated from corrupted R01s. Further, a "best" extrapolation method for all texture descriptors was not identified, which implies that the choice of extrapolation method depends on the texture descriptors applied in a given tissue classification scheme.