OA145] Impact of tracer activity reduction on PET radiomic features in pediatric 18F-FDG-PET/MRI examinations

Abstract

Purpose The evaluation of features robustness in respect to image acquisition and post-processing parameters changes is fundamental for radiomic studies, especially in the case of multi-institutions collaborations. In this study the sensibility of PET radiomic features to tracer activity reduction in PET/MRI pediatric examinations was investigated in detail. The use of different SUV resampling and contouring strategies was also assessed. Methods Twenty-one whole-body 18F-FDG-PET/MRI pediatric examinations performed on a clinical PET/MR scanner (Siemens Biograph mMR) were considered. To simulate activity reduction, the full tracer activity (FTA) list-mode data (3 MBq/kg, 5 min/bed position) were truncated corresponding to decreasing activities (2.4, 1.8, 1.5, 1.2 and 0.6 MBq/kg). Fifty-six lesions were detected. Two methods of VOI definition have been employed: 1) using a threshold equal to 40% of SUVmax in the FTA image and copying the identified VOIs to low-activity images (Fixed VOIs); 2) thresholding at 40% of SUVmax in each image (Variable VOIs). To reduce noise, images were pre-processed discretizing lesion SUV in two different ways: 1) resampling to 64 levels (R64); 2) rounding to 0.25 (R025). Fifty-five texture features were extracted using IBEX (v.1.0 β ) . To assess features robustness, for each metric and reduced activity level, relative changes from FTA image, paired Wilcoxon tests and Intraclass Correlation Coefficient (ICC) were computed taking FTA features values as the ground truth. SUVmean was computed in the lesions at each activity level and its average relative changes from FTA were considered as benchmarks for texture features stability. Results The choice of contouring and SUV resampling methods results in different features changes at lower activity levels from FTA. Focusing on Variable VOI and R025, at 1.5 MBq/kg, 35/55 features showed both non-statistically significant differences from FTA image (paired Wilcoxon test, p-value 0.01) and ICC > 0.85. At lowest activities (1.5, 1.2, 0.6 MBq/kg), about half of the features changed from FTA values less than the corresponding SUVmean changes (2%, 3%, 9%) using Variable VOI-R025. Conclusions Specific PET texture features can be reliably extracted from low-activity images in pediatrics for tumor biologic heterogeneity characterization depending on tracer activity level, SUV resampling and lesion delineation strategies.