Program Information
Impact of Radial Location Inside a CT QA Phantom On Radiomic Features Using 3D-Printed Cylindrical Textured Inserts
M Shafiq ul Hassan1,2*, G Zhang1,2 , K Latifi1,2 , G Ullah2 , D Hunt1 , R Gillies1 , E Moros1,2 , (1) H Lee Moffitt Cancer Center and Research Institute, Tampa, FL, (2) University of South Florida, Tampa, FL,
Presentations
WE-RAM2-GePD-IT-2 (Wednesday, August 2, 2017) 10:00 AM - 10:30 AM Room: Imaging ePoster Theater
Purpose: Computed Tomography (CT) image reconstruction is affected by the relative differences in the variation of photon energy and numbers between the center and other radial locations of the phantom/patient. Here, we compared radiomic features extracted from region of interests (ROI) located at center and periphery of a commonly used CT quality assurance (QA) phantom using 3D printed textured inserts.
Methods: CT scans of two identical 3D printed cylindrical textured inserts within a Gammex 467 (G467) phantom were acquired using GE Discovery and Philips 64 scanners. The phantom was aligned such that one insert was positioned along the longitudinal axis of the scanner and another at the periphery of the G467. Moreover, the periphery position of the insert was changed to different angles from 45 to 360° for 8 scans per scanner. A spherical ROI was contoured on the images of the textured inserts. Radiomic features were extracted using an in-house program. Feature categories were: Shape (11), Intensity (16), GLCM (26), GLSZM (11), GLRLM (11), and NGTDM (5) for a total of 80 features. The absolute percent difference (%Diff.) between center and periphery was calculated for all features. % Diff. results were arbitrarily classified into three groups: negligible %Diff. < 10, small 10 ≤ %Diff ≤ 30 and large %Diff. > 30 differences.
Results: Shape features were robust for both scanners. Intensity features were sensitive to scanner, and 44% and 13% of features had %Diff. ≥ 30 for GE and Philips scanner, respectively. GLCM and GLRLM features had, respectively 21% and 32% of features with %Diff. > 30. Most of GLSZM and some NGTDM features showed %Diff. >30.
Conclusion: Almost 35-40% of all 53 texture features were sensitive to the position of ROI within the phantom image. This may be due to the correlated noise intrinsic to image reconstruction.
Funding Support, Disclosures, and Conflict of Interest: Funded in part by Grant NIH/NCI RO1CA190105-01
Contact Email: