Program Information
Normalization Effects On CT Texture Feature Stability
I Mihaylov1*, (1) Univ Miami, Miami, FL
Presentations
WE-RAM1-GePD-J(A)-2 (Wednesday, August 2, 2017) 9:30 AM - 10:00 AM Room: Joint Imaging-Therapy ePoster Lounge - A
Purpose: To explore the effects of feature normalization on the magnitude of feature variability in gray-level co-occurrence matrices (GLCMs).
Methods: A Cathphan phantom was CT scanned seven times over a period of three weeks on Siemens 64-slice scanner. The reconstructed CT images had a resolution of 0.976x0.976x1 mm3. Four regions of interest (ROIs) were manually outlined on the phantom for each scan. Two of those regions had fairly uniform HU values, while the other two had rather variable HU. For each ROI four GLCMs were created – with 64 and 32 bins, and with fixed bin width of 1 and 4 HUs, covering the range of HUs specific for each particular ROI. Eighteen commonly used image features were calculated form the GLCMs, and the variability (defined as standard deviation divided by the average) of those features were calculated for all GLCMs. Variability was scaled with ROI volume and range and was evaluated with respect to the unscaled variability for each GLCM.
Results: For the two ROIs with uniform density the variability of the mean, variance, and energy decreased when the corresponding quantities were multiplied by the RIO volumes. Variability of the majority of the remaining features for those ROIs were also reduced when the individual features were normalized to the range of HUs or to the ROI volumes. For non-uniform density 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 range of the ROIs.
Conclusion: GLCM feature variability depends on volume and range. The dependence is variable and is affected by the binning approach used to generate the GLCMs. Fixed bin width GLCM variability is weakly dependent on the tested normalization factors.
Contact Email: