Click here to


Are you sure ?

Yes, do it No, cancel

Homogenizing Reconstruction Kernels for CT Radiomics

D Mackin1*, R Ger2 , L Zhang3 , J Yang4 , P Chi5 , S Bache6 , C Dodge7 , A Jones8 , L Court9 , (1) MD Anderson Cancer Center, Houston, TX, (2) MDACC, Houston, TX, (3) UT MD Anderson Cancer Center, Houston, TX, (4) UT MD Anderson Cancer Center, Houston, TX, (5) M.D. Anderson Cancer Center, Houston, TX, (6) Mission Health System, Asheville, NC, (7) Texas Children's Hospital, Houston, TX, (8) UT MD Anderson Cancer Center, Houston, TX, (9) UT MD Anderson Cancer Center, Houston, TX


(Tuesday, 7/31/2018) 1:45 PM - 3:45 PM

Room: Karl Dean Ballroom C

Purpose: Quantitative feature values depend on the convolution kernels used to reconstruct the images. The purpose of this study was to determine the most compatible kernels for the leading CT scanner manufacturers and to investigate a new image filtering method to homogenize images reconstructed using discordant kernels.

Methods: A CT radiomics phantom was imaged using scanners from GE, Philips, Siemens, and Toshiba. The images from each scanner were reconstructed multiple times while varying the reconstruction kernels from smooth to sharp, with 6-8 levels from each manufacturer. The phantom comprised 10 cartridges of materials producing a range of image textures. For each scan, the materials were semi-automatically contoured to produce 8×2×2cm3 regions-of-interest, and 38 quantitative features were extracted from the categories first order intensity (n=12), neighborhood gray-tone difference matrix(n=5), and gray-level co-occurrence matrix(n=21). The fractional differences between the features for all kernels and those from the chosen baseline kernel (GE Standard) were calculated. To gauge the effect size, these differences were scaled by the coefficient-of-variation (CV) of the same features extracted from a cohort of non-small cell lung cancer patients. Using the noise power spectrum for each kernel, we developed filters to make the image frequency magnitudes similar to those of the baseline kernel.

Results: The Philips C, Siemens B30f, and Toshiba FC08 kernels were most similar to the baseline. Feature values strongly depended on the kernels--the range of the feature entropy was 2.8 times the patient CV. Of the kernels tested, 17 of 27 were incompatible (median difference from the baseline was >0.5 × patient CV). After applying kernel-specific filters, all GE, Philips, and Siemens kernels were compatible except for GE Edge.

Conclusion: For retrospective studies, the effects due to discordant kernels can be mitigated by applying kernel-specific frequency space image filters.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email