237. Robustness of FDG-PET radiomic features against tumor delineation in two clinical relevant scenarios

Purpose To assess the robustness of PET radiomic features (RF) against tumour/PET positive nodes (T/N) delineation uncertainty in two clinically relevant situations, namely head-and-neck (HNC) and pancreatic cancer. Methods Twenty-five HNC (T and N) and 25 pancreatic cancer (only T) patients previously treated with IMRT and 18 FDG-PET/CT based planning optimization were included in the study. Seven contours were delineated for each lesion following manual (2 independent observers), semi-automatic (based on SUV_max gradient: PET_edge) and automatic (40%,50%,60%,70% SUV_max thresholds) methods using MIM software(MimVista). For each contour, 71 RF (12 of first and 59 of higher order) were extracted by using the CGITA (v1.4) software. The impact of delineation on volume agreement (DICE-index) and RF was assessed by Intra-class Correlation Coefficients (ICC), considering a threshold of 0.80. Semi-automatic contouring repeatability was also tested. Results A large disagreement between manual and automatic SUV_max method was found for threshold values ⩾50%. Inter-observer variability was moderately high with median DICE values between 0.81 (HN-T) and 0.73 (pancreas). Volumes defined by PET_Edge were in better agreement with the manual ones compared to SUV40%; the repeatability of PET_Edge was excellent (median DICE = 0.95). Regarding RF, 14/14/35 features showed an ICC  Conclusions About 80%/50% of 72 RF extracted with the CGITA software were consistent with respect to manual delineation for HN/pancreas patients. PET_edge was sufficiently robust with about 60%/40% of RF consistent with manual delineation while a worse performance was found for SUV40%. This result suggests the possibility to replace manual delineation of HN and pancreas tumours in studies including PET radiomic analyses.