First Step Towards Radiomics In Ga-68-Dotatate Pet/Ct: Feature Reproducibility

647 Objectives: Well-differentiated neuroendocrine tumors (NET) are characterized by overexpression of somatostatin receptors. Radiolabeled somatostatin analogues specifically target these receptors, enabling diagnostic imaging and peptide receptor radionuclide therapy (PRRT) in inoperable and/or metastasized NET. The phase III NETTER-1 trial shows an overall increase in progression free survival after PRRT, though actual response varies among patients. PRRT is generally expected to be more effective in lesions with high uptake on diagnostic somatostatin analogue imaging like Gallium-68 (68Ga) DOTATATE PET/CT. By applying image analysis and identifying features that are tumor-specific (also known as radiomics), prediction of treatment response has proven feasible in lung and head-neck neoplasms. The objective of this initial study was to evaluate the feasibility of radiomics in NET through evaluation of feature reproducibility in 68Ga-DOTATATE PET/CT. Methods The NEMA IQ phantom was filled with several clinically relevant tumor-to-background concentrations of 68Ga-DOTATATE. Accordingly, the minimal diameter to quantify target lesions was determined using recovery coefficient curves. A cohort of 30 patients with histology proven advanced well-differentiated NET in whom sequential 68Ga-DOTATATE PET/CT scans were acquired two days apart was available to determine feature reproducibility (test-retest setting). Whole body PET-acquisitions were made according to local clinical protocol 45 minutes after intravenous injection of approximately 100 MBq. All scans were acquired on a Philips Gemini ToF BigBore at 2.5 min/bed-position (iterative Blob-OS-TF reconstruction, 4×4×4mm voxel size). Target lesions were delineated on both scans with the open source software platform 3DSlicer v4.8 using the GrowCut algorithm and manually adjusted by one observer. For this pilot study, 19 first-order intensity statistics and 16 shape features were calculated using the PyRadiomics Toolbox. Concordance of the shape features between the sequential PET-scans was used to compare delineations. Statistical analysis included the intraclass correlation coefficient (ICC) to evaluate feature stability and coefficient of variation (CoV) evaluates absolute agreement between both measurements. Based on ICC reproducibility was classified high (>0.85), moderate (0.5-0.85) or poor (