PET tumor segmentation: Validation of a gradient-based method using a NSCLC PET phantom

Objectives Reproducible, accurate PET segmentation methods are needed to calculate statistics for assessing therapy response and to aid in creation of target volumes for Radiation Oncology. Previously, we compared a gradient-based method, PET Edge, (GRAD) to constant threshold (THRESH) using spherical phantoms. Our goal is to evaluate the accuracy of GRAD and THRESH on realistic, simulated PET phantom data for target volume definition and quantitative assessment of tumor activity.

Methods Twenty-five realistic digital PET phantoms of the thorax were obtained with 31 simulated tumors of varying size, shape and location. An observer segmented each tumor with GRAD and THRESH. THRESH was performed using thresholds of 15-50% of maximum counts at 5% increments. Tumor volumes for each method were compared to known volumes from digital phantoms. Total Glycolytic activity (TGA), SUVmean * volume, was calculated for each method. Tumors were grouped by size into <60ml, 60-120ml, and >120ml. Mean absolute % difference was calculated for the volume (Vdiff%) and TGA (TGAdiff%) for each group using all methods.

Results GRAD achieved greater accuracy than any THRESH method. For tumors <60ml, 60-120ml, and >120ml, the Vdiff% using GRAD was 14.2%, 8.0%, and 6.2%, respectively. Vdiff% for the best THRESH was 35.8%, 27.3%, and 18.6%, respectively. TGAdiff% for GRAD was 7.7%, 3.3%, and 1.5%, respectively. TGAdiff% for the best THRESH was 27.0%, 13.8%, and 15.7%, respectively.

Conclusions GRAD resulted in more accurate tumor volumes and TGA statistics than THRESH for all lesions. Through more accurate tumor volume segmentation and statistics, GRAD may play an important role in prognosis, assessing therapy response, and creating target volumes in Radiation Oncology.