Reproducibility of categorical metabolic response assessment using a novel semi-automatic software versus a manual analytical method

Objectives: To assess the reproducibility of metabolic tumor response based on EORTC and PERCIST 1.0 criteria using a novel semi-automated software [HERMES Hybrid3D-Tumor Finder (3D)] compared with a manual FDA-approved platform [HERMES Hybrid Viewer 2.2D (2D)].

Methods: 13 patients with GIST enrolled in a prospective clinical trial testing a novel drug had baseline and week 8 FDG-PET/CT scans. One imaging analyst identified up to 5 target lesions at both time points, manually drew volumes of interest (VOIs) to determine SUVmax and SUVpeak using a 2D method, then re-analyzed the datasets using 3D. Categorical metabolic response assessments of Partial Metabolic Response (PMR), Stable Metabolic Disease (SMD), and Progressive Metabolic Disease (PMD) were performed according to EORTC and PERCIST criteria.

Results: For PERCIST: 12 patients had identical response categories assigned with both 2D and 3D (n=1 PMR, n=9 SMD, n=2 PMD). One patient was excluded because there was no measureable disease detected at baseline using the 3D semi-automated method preventing a response assessment while the 2D method determined SMD based on the manual placement of a VOI at both time points. The single hottest lesion chosen by the two methods differed in 1 patient at baseline and at week 8. Differences in the choice of the PERCIST target lesion (one hottest SUVpeak) were due to differences in the size of VOIs drawn and the semi-automatic background voxel exclusion constraints imposed by the 2D vs 3D methods. For exploratory PERCIST analysis of the hottest 5 SUVpeak lesions, 3D identified lesions in 8 patients that were not identified by the image analyst on 2D. For EORTC: 13 patients had identical response categories assigned with both methods (n=1 PMR, n=9 SMD, n=3 PMD). The target lesions (5 hottest SUVmax) chosen by the two methods differed in 10/13 patients. For the same lesions, similar SUVmax measurements were recorded at each time point (72 lesions = 0% difference, 12 lesions &lt 1%, 3 lesions &lt 10% , 4 lesions &lt 20% , 1 lesion &lt 36%).

Conclusion: While the choice of target lesions may differ between the two HERMES software methods (2D vs 3D), identical metabolic response categories were assigned for both PERCIST and EORTC criteria. Both software are therefore equally relevant clinically and the semi-automated method allows for better ease of lesion identification and measurement, as well as an improved use of time.