Improved Stability and Performance of ¹⁸F-FDG PET Automated Tumor Segmentation Using Multi-level Maximization of Inter-class Variance Method

Objectives There is an emerging need for a stable and consistent method for volumetric delineation of tumor PET data as current methods can produce varying results depending on the tumor sizes and FDG uptake levels. We developed a novel automated threshold method and tested its accuracy across different tumor sizes and PET standardized uptake values (SUVs) compared to current methods.

Methods Twenty-five tumors of varying sizes (range: 0.59 to 21.01 cm³) and SUVs (range: 3.17 to 25.80) on FDG PET were segmented using a multi-Otsu based (MO) method, which maximized inter-class variance, to derive metabolic tumor volume (MTV). Results were compared with MTVs from 7 different PET threshold methods: 20%, 40%, 60%, or 80% of the maximum SUV (SUV_max), and mean liver SUV + 1, +2, and +3 standard deviations (SD). CT-based gross tumor volume (GTV) was used as a reference standard. To evaluate segmentation accuracy, we compared volume ratio (VR) or the ratio of MTV to GTV. VR closer to 1 pointed toward a better segmentation strategy.

Results MO method, among all tested methods, resulted in the most consistent mean VR (closest to 1) for a wide range of tumor sizes: 0 to 1 cm³ tumors (mean VR ± SD = 1.33 ± 0.32), 1 to 2 cm³ tumors (mean VR ± SD = 1.10 ± 0.41), 2 to 3 cm³ tumors (mean VR ± SD = 1.09 ± 0.22), and tumors greater than 3 cm³ (mean VR ± SD = 1.06 ± 0.17). MO method also consistently produced VR that is closest to 1 for tumors with varying PET SUVs or FDG uptake levels: SUV_max 0 to 5 (mean VR ± SD = 1.32 ± 0.32), SUV_max 5 to 10 (mean VR ± SD = 1.16 ± 0.30), SUV_max 10 to 15 (mean VR ± SD = 1.00 ± 0.19), and SUV_max > 15 (mean VR ± SD = 1.09 ± 0.10).

Conclusions We have demonstrated a novel PET tumor segmentation method, which allows for stable and consistent delineation across a range of clinically applicable tumor sizes and PET SUV values compared to current methods.