TY - JOUR
AR - RDI-2018-1-105
TI - Computed tomography texture feature stability dependence on the graylevel co-occurrence matrix discretization approach
AU - Ivaylo, Mihaylov
JO - Radiology and Medical Diagnostic Imaging
PY - 2018
DA - Fri 28, Sep 2018
SN - 2613-7836
DO - http://dx.doi.org/10.31487/j.RDI.2018.10.005
UR - https://www.sciencerepository.org/ct-textural-feature-stability-and-discretization-approach_RDI-1-105 
KW - quantitative, imaging, biomarker, stability, CT, GLCM
AB - Purpose: To explore the effects of the discretization approach on the stability of the second-order computed tomography (CT) based textural features, derived from gray-level co-occurrence matrices (GLCMs). 
Material and Methods: A Cathphan phantom was scanned seven times over three weeks. Four cylindrical regions (ROIs) were manually outlined for each scan. Two regions had fairly uniform density, while the other two were more heterogeneous. For each ROI four GLCMs were created – with 64 bins, with 32 bins, and with fixed bin widths of 1 and 4 HU. Eighteen commonly used radiomics features were calculated form the GLCMs, and their variabilities were compared among the four GLCM representations.
Results: The uniform ROIs had average standard deviation of the HUs of ~1.5%, while the heterogeneous ROIs had standard deviations greater than 4%. For the uniform ROIs the variability of the fixed number of bins GLCMs was on average lower than the variability for fixed bin width GLCMs. For the heterogeneous ROIs the situation was reversed. For the uniform ROIs the variability of the mean, variance, and energy decreased when the corresponding quantities were multiplied by the RIO volumes. Variabilities of the majority of the remaining features for those ROIs were also reduced when the features were normalized to the HU ranges or to the ROI volumes. For heterogeneous ROIs the mean, variance, energy, auto correlation, and correlation were weakly dependent on volume and range. The variability of fixed number of bins GLCMs exhibited strong dependence on the ROI range.
Conclusion: This study indicates that the GLCMs creation is affected differently depending on the homogeneity of the ROI. The fixed number of bins GLCMs produce fewer variable features for homogenous objects and vice-versa. Additionally, it was demonstrated that for realistic patient scenarios the use of fixed bin width GLCMs may be advantageous.