Evaluation of threshold and gradient based (18)F-fluoro-deoxy-2-glucose hybrid positron emission tomographic image segmentation methods for liver tumor delineation.

PURPOSE

Image segmentation methods were studied to delineate liver lesions in (18)F-fluoro-2-deoxy-glucose positron emission tomographic (FDG-PET) images. The goal of this study was to identify a clinically practical, semiautomated FDG-PET avid volume segmentation method to improve the accuracy of liver tumor contouring for treatment planning in stereotactic body radiation therapy (SBRT).

METHODS AND MATERIALS

Pretreatment PET-CT image sets for 26 patients who received SBRT to 28 liver lesions were delineated using the following 3 methods: (1) Percent threshold with respect to background corrected maximum standard uptake values (SUV; threshold values varied from 10% to 50% with 10% increments); (2) threshold 3 standard deviations above mean background SUV (3σ); and (3) a gradient-based method that detects the edge of the FDG-PET avid lesion (edge). For each lesion, semiautomatically generated contours were evaluated with respect to reference contours manually drawn by 3 radiation oncologists. Two similarity metrics, Dice coefficient, and mean minimal distance (MMD), were employed to assess the volumetric overlap and the mean Euclidian distance between semiautomatically and observer-drawn contours.

RESULTS

Mean Dice and MMD values for 10%, 20%, 30% threshold, 3σ, and edge varied from 0.69 to 0.73, and from 3.44 mm to 3.94 mm, respectively (ideal Dice and MMD values are 1 and 0 mm, respectively). A statistically significant difference was not observed among 10%, 20%, 30% threshold, 3σ, and edge methods, whereas 40% and 50% methods had inferior Dice and MMD values.

CONCLUSIONS

Three PET segmentation methods were identified above as potential tools to accelerate liver lesion delineation. The edge method appears to be the most practical for clinical implementation as it does not require calculation of SUV statistics. However, the performance of all segmentation methods showed large lesion-to-lesion fluctuations. Therefore, such methods may be suitable for generating initial estimates of FDG-PET avid volumes rather than being surrogates for manual volume delineation.