Abstract
2077
Objectives Assessing response to therapy or prognosis using SUVmax has a number of limitations including sensitivity to noise. Total lesion glycolysis (TLG) is less sensitive to noise and incorporates both size and activity (SUVmean x volume) into one measure. Accuracy of PET segmentation is important for TLG since it utilizes volume. Previously we demonstrated the superior volumetric accuracy of a gradient PET segmentation method (GRAD) compared to thresholds (THRESH) and manual contouring (MC) in realistic Monte Carlo simulated PET scans of the thorax. In this study we evaluated the effect of segmentation accuracy on TLG.
Methods Thirty-one lung tumors of varying size, shape, and location were segmented by 7 clinicians on 25 realistic digital PET scans of the thorax. GRAD, THRESH and MC methods were used. GRAD identifies tumor edges based on a change in count levels at the tumor border. THRESH was performed using 25-50% of maximum counts at 5% increments. Accuracy and bias were measured by calculating the mean absolute % error and mean % errors respectively (abs%error and %error) for TLG using all methods.
Results GRAD was the most accurate technique with abs%error of 6.1 (10.3 SD). Both 25% THRESH, the most accurate threshold, and MC were significantly less accurate with abs%error of 10.2 (19.1) and 14.5 (17.2) respectively (p < 0.0013). 25% THRESH had the smallest bias with %error of -1.1 (21.6) followed by GRAD with -3.8 (11.4), however, the difference was not significant (p = 0.13).
Conclusions GRAD resulted in significantly more accurate TLG calculations than the other contouring methods. GRAD has the potential to play an important role in both determining prognosis and assessing response to therap