Abstract
1405
Objectives A PET tumor segmentation method using multi-level Otsu (MO) has been developed and shown stable and consistent results across a wide range of tumor sizes and tumor uptake in clinical FDG PET/CT scans. We further validated its accuracy using standard NEMA image quality (IQ) phantom compared to current methods and tested in clinical cases.
Methods The NEMA IQ phantom was filled with an 18F solution to have a uniform background activity. The six spherical lesions in the phantom (volume: 0.52, 1.15, 2.57, 5.57, 11.5, and 26.5 cm3) were filled with 18F activity to have a lesion-to-background ratio (LBR) of either 8:1, 4:1, or 1.5:1. The phantom was imaged using a GE Discovery 710 PET/CT scanner. The lesions in the phantom were segmented using our MO method, to derive a spherical lesion metabolic tumor volume (MTV). Results were compared with MTVs from 8 different PET threshold methods: 20%, 40%, 60%, or 80% of the maximum activity (Bq/ml), and mean background + 1 or +2 standard deviations (SD), and mean background x 2 or x 2.5. Three small lesions were not evaluated in LBR of 1.5:1, because they were not distinguishable from background at this low LBR. To evaluate PET segmentation accuracy, we compared volume ratio (VR) of MTV to the actual volume of the phantom lesions. VR closer to 1 indicated a better segmentation strategy. We further tested MO method in clinical cases of osteosarcoma and prostate cancer.
Results MO method and 40% threshold showed more consistent mean VRs (closest to 1) for each combination of different lesion sizes and LBR ratios than the other methods. Our MO method showed stable and relatively accurate estimation of the true volume in all lesions and LBRs. However, 40% threshold substantially overestimated MTV in small lesions or in the setting of low LBR, in contrast to our MO method (Figure 1). Table 1 shows representative VR data for the MO method, 40%, and mean background +2 standard deviations (BG + 2SD). MO method could be successfully applied for tumor segmentation in clinical cases with heterogenous uptake or multiple lesions (Figure 2, Figure 3).
Conclusions We have validated in NEMA IQ phantom studies that our MO PET segmentation method is relatively accurate, and importantly, stable and consistent across a range of lesion sizes and LBRs representative of clinical tumor lesions. This segmentation method holds promise for clinical PET segmentation applications but requires further validation in clinical PET data sets.