Encrypted login | home

Program Information

Assessment of Reproducibility Of FDG-PET-Based Radiomics Features Across Scanners Using Phantom Imaging


D Fried

D Fried1*, J Meier2 , O Mawlawi2 , S Zhou2 , G Ibbott2 , Z Liao2 , L Court2 , (1) University of North Carolina at Chapel Hill, Chapel Hill, NC, (2) UT MD Anderson Cancer Center, Houston, TX

Presentations

MO-DE-207B-7 (Monday, August 1, 2016) 1:45 PM - 3:45 PM Room: 207B


Purpose: Use a NEMA-IEC PET phantom to assess the robustness of FDG-PET-based radiomics features to changes in reconstruction parameters across different scanners.
Methods: We scanned a NEMA-IEC PET phantom on 3 different scanners (GE Discovery VCT, GE Discovery 710, and Siemens mCT) using a FDG source-to-background ratio of 10:1. Images were retrospectively reconstructed using different iterations (2-3), subsets (21-24), Gaussian filter widths (2, 4, 6mm), and matrix sizes (128,192,256). The 710 and mCT used time-of-flight and point-spread-functions in reconstruction. The axial-image through the center of the 6 active spheres was used for analysis. A region-of-interest containing all spheres was able to simulate a heterogeneous lesion due to partial volume effects. Maximum voxel deviations from all retrospectively reconstructed images (18 per scanner) was compared to our standard clinical protocol. PET Images from 195 non-small cell lung cancer patients were used to compare feature variation. The ratio of a feature’s standard deviation from the patient cohort versus the phantom images was calculated to assess for feature robustness.
Results: Across all images, the percentage of voxels differing by <1SUV and <2SUV ranged from 61-92% and 88-99%, respectively. Voxel-voxel similarity decreased when using higher resolution image matrices (192/256 versus 128) and was comparable across scanners. Taking the ratio of patient and phantom feature standard deviation was able to identify features that were not robust to changes in reconstruction parameters (e.g. co-occurrence correlation). Metrics found to be reasonably robust (standard deviation ratios > 3) were observed for routinely used SUV metrics (e.g. SUVmean and SUVmax) as well as some radiomics features (e.g. co-occurrence contrast, co-occurrence energy, standard deviation, and uniformity). Similar standard deviation ratios were observed across scanners.
Conclusions: Our method enabled a comparison of feature variability across scanners and was able to identify features that were not robust to changes in reconstruction parameters.




Contact Email: