Click here to

Session: Imaging General ePoster Viewing [Return to Session]

Assessing the Variability of Quantitative Imaging Features Across a Range of CT Acquisition and Reconstruction Parameters: Investigating Whether Data From One Scanner Can Be Used to Represent the Range of Data From Other Scanners

G Melendez-Corres*, G Kim, M Wahi-Anwar, M McNitt-Gray, David Geffen School of Medicine at UCLA, Los Angeles, CA


PO-GePV-I-60 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Machine learning algorithms or quantitative imaging systems should be tested across a range of acquisition and reconstruction parameters to reflect clinical realities. The purpose of this work was to assess the variability in radiomic feature values between scanners from different manufacturers for the potential of building datasets capable of covering the variability of other manufacturer’s scanners when scanners from a single manufacturer are used.

Methods: Image data was collected from a public dataset of scans of the Credence Cartridge Radiomics (CCR) phantom. In this dataset, scans were performed on Siemens, GE and Philips scanners. Three subsets of this data were chosen for analysis where kernel (n=28), mAs (n=20), and slice thickness (n=24) were varied one parameter at a time. A 2D circular ROI was placed on the centermost slice of the glued rubber cartridge section of the CCR phantom. From this ROI, four first order and four GLCM texture features were calculated. The Coefficient of Variation (CV) was computed for each feature vector from Siemens, GE, and Philips to compare their variability. To compare whether Siemens features were different from GE/Philips, Student’s T-test, Mann-Whitney U test and Wilcoxon signed-rank test were performed as appropriate for each feature vector.

Results: Across all three variations and eight radiomic features, the CV from Siemens was higher than from GE on 23/24 combinations and higher than Philips on 18/24 combinations. Siemens was statistically different from GE/Philips on 4 features: mean value on the kernel variations, uniformity and entropy value on the mAs variations, and mean value on the slice thickness variations.

Conclusion: This study gives insight on the possibility of generating robust data that covers variability from other manufacturers when scanners from only one manufacturer are available. Some additional work is needed to address sources of variability due to heterogeneity within the phantom.

Funding Support, Disclosures, and Conflict of Interest: M McNitt-Gray: Grant funding through the Master Research Agreement between UCLA and Siemens Healthineers. G Kim: MedQIA, consultant


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email