Program Information
Impact of Algorithmic Implementation On Quantitative Texture Feature Values
J Foy*, K Mendel , H Li , M Giger , H Al-Hallaq , S Armato , The University of Chicago, Chicago, IL
Presentations
TH-AB-201-4 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: 201
Purpose: Several open-source texture analysis software packages have recently been developed. This study reports on the variation in the texture calculation algorithms across four such packages. Knowledge of such potential variations is crucial as quantitative image analysis (radiomics) is translated to precision medicine.
Methods: Four texture analysis packages were used: two developed within our research labs, MaZda, and IBEX. A total of 40 265x265-pixel ROIs from digital mammography images of 38 healthy patients were used as input to calculate six first-order features (maximum, minimum, mean, standard deviation, skewness, kurtosis) and four second-order gray-level co-occurrence matrix (GLCM) features (contrast, entropy, difference entropy, sum average). Features were first calculated using the default GLCM parameters and then modified in pixel spacing, gray-level limits, and number of directions to provide greater consistency among packages. Features are reported as the mean and standard deviation across all 40 ROIs. Non-parametric tests were used to compare pairs of feature values using IBEX as the control, and significance was assessed at p<0.001 to correct for multiple comparisons.
Results: First-order features agreed to within 1.13%±0.10% except for kurtosis, which varied by 101%±20%. Kurtosis was the only feature to show statistically significant differences between the packages. The means of all second–order features differed significantly between all four packages by up to four orders of magnitude. When the GLCM parameters were modified, the feature values still differed by as much as one order of magnitude, but still showed statistically significant differences between packages for all features. These differences could mostly be attributed to the variation in the parameters used to generate the GLCMs.
Conclusion: The large variations in the values of the second-order texture features indicate that texture analysis packages should be tailored to the images being analyzed with special attention to the algorithmic settings.
Contact Email: