RT Journal Article SR Electronic T1 Repeatability of 18F-FDG PET radiomic features in pelvic malignancies JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 272 OP 272 VO 61 IS supplement 1 A1 John Crandall A1 Tyler Fraum A1 Min-Young Lee A1 Perry Grigsby A1 Richard Wahl YR 2020 UL http://jnm.snmjournals.org/content/61/supplement_1/272.abstract AB 272Introduction: An understanding of test-retest variability is essential for suitable interpretation of radiomic features. The primary aim of this study was to assess the repeatability of radiomic features extracted from 18F-Fluorodeoxyglucose (FDG) positron emission tomography (PET) images of pelvic malignancies. The impact of spatial resampling and standardized uptake value (SUV) discretization on feature repeatability was also explored. Methods: Eleven female patients with pathology-proven pelvic malignancies were evaluable for this prospective study. Baseline and repeat FDG PET/CT imaging were performed within 7 days, with no oncologic treatment between imaging sessions. The same PET/CT scanner and approximate FDG dose were used for both imaging sessions per the Quantitative Imaging Biomarkers Alliance (QIBA) FDG PET profile. Pelvic tumors were segmented to produce whole tumor volumes of interest (VOIWT) and 40% isocontours of the whole tumor VOIs (VOI40). Tumors were segmented using MIM 6.9 and 44 radiomic features were extracted using LIFEx 5.28. Radiomic features were extracted from both VOIs using various extraction parameters. Voxel size was either left as the default scanner size or resampled to 3 mm isotropic voxels. SUV was discretized using 32, 64, or 128 bins, which covered the range of SUV values in the tumor VOI. Lin's concordance correlation coefficient (CCC) was used to assess the repeatability of extracted radiomic features, with a CCC value of at least 0.80 considered to be repeatable. Results: Eleven patients were enrolled and underwent double baseline FDG PET/CT imaging. 41% (18/44) of radiomic features were repeatable (CCC > 0.80), regardless of extraction parameters. Radiomic features extracted from VOI40 showed significantly better repeatability than features extracted from VOIWT (P < 10-27). For most features (78.4%), a change in bin number or voxel size resulted in less than 10% change in the feature value. Shape features (volume, sphericity, and compacity) were repeatable (CCC values > 0.82), regardless of extraction parameters or segmentation type, as were most gray level co-occurence matrix (GLCM) features (5/6) and neighborhood gray level difference matrix (NGLDM) features (2/3). All gray-level emphasis and gray-level run emphasis features showed poor repeatability (CCC values < 0.52) when extracted from ROIWT, but were highly repeatable (CCC values > 0.96) when extracted from ROI40. Conclusions: Certain groups of radiomic features (shape, GLCM, and NGLDM) were consistently repeatable, while other groups (gray level run length matrix and gray level zone length matrix features) were highly variable. Radiomic features extracted from 40% isocontours were more repeatable than features extracted from whole tumor segmentations. Most often, changes in voxel size or SUV discretization parameters resulted in relatively small differences in feature value, though several features were highly sensitive to these changes. Acknowledgment: This project was supported by NIH grant U01 CA140204.