Threshold segmentation for PET target volume delineation in radiation treatment planning: The role of target‐to‐background ratio and target size

Abstract

A multivariable approach was adopted to study the dependence of the percentage threshold [TH(%)] used to define the boundaries of 18F-FDG positive tissue on emission scan duration (ESD) and activity at the start of acquisition (Aacq) for different target sizes and target-to-background (T/B) ratios. An anthropomorphic model, at least for counting rate characteristics, was used to study this dependence in conditions resembling the ones that can be encountered in the clinical studies. An annular ring of water bags of 3 cm thickness was fitted over an International Electro-technical Commission (IEC) phantom in order to obtain counting rates similar to those found in average patients. The scatter fraction of the modified IEC phantom was similar to the mean scatter fraction measured on patients, with a similar scanner. A supplemental set of microhollow spheres was positioned inside the phantom. The NEMA NU 2-2001 scatter phantom was positioned at the end of the IEC phantom to approximate the clinical situation of having activity that extends beyond the scanner field of view. The phantoms were filled with a solution of water and 18F (12 kBq/mL) and the spheres with various T/B ratios of 22.5, 10.3, and 3.6. Sequential imaging was performed to acquire PET images with varying background activity concentrations of about 12, 9, 6.4, 5.3, and 3.1 kBq/mL. The ESD was set to 60, 120, 180, and 240 s/bed. Data were fitted using two distinct multiple linear regression models for sphere ID < or = 10 mm and sphere ID > 10 mm. The fittings of both models were good with an R2 of 0.86 in both cases. Neither ESD nor Aacq resulted as significant predictors of the TH(%). For sphere ID < or =10 mm the target size was the most significant predictor of the TH(%), followed by the T/B ratio, while for sphere ID > 10 mm the explanatory power of the target size and T/B ratio were reversed, the T/B ratio being now the most important predictor of the TH(%). Both the target size and T/B ratio play a major role in explaining the variance of the TH(%), throughout the whole range of target sizes and T/B ratios examined. Thus, algorithms aimed at automatic threshold segmentation should incorporate both variables with a relative weight which critically depends on target size.

Adaptive threshold segmentation of pituitary adenomas from FDG PET images for radiosurgery

“…This led to a larger threshold being required to the segment smaller volumes than what is visually seen. This is similar to the relationship observed and reported by others 20 , 21 …”

“…The tumor size contributes less to the threshold prediction when compared to T/B ratio $false(\beta =0.15false)$, but still significantly. This trend is similar to that reported by Brambilla et al (21) and Drever et al (29) …”

“…In our own experiments we found it to be relevant to the tumors in question. Brambilla et al (21) also found similar relation and concluded that parameters like injected activity and scan durations had a negligible role in predicting an optimal threshold (18) . Based on suggestions from their work for the clinical implementation, we used the activity ratio of target to its background as measured from the images and not the actual T/B as used in the phantom.…”

“…The transformation (1‐B/T), as suggested by Brambilla et al, (21) and the inverse of size of the target expressed in cm 3 showed a linear relationship with the baseline threshold. So, a linear model with these two variables of the form $\mathrm{T}=\mathrm{k}+\mathrm{a}\ast false(1-\mathrm{B}/\mathrm{T}false)+\mathrm{b}\ast false(1/\text{size}false)$ was assumed as the predictor equation for threshold, where k, a, b are the regression coefficients.…”

“…As the equation requires the volume of the target to compute the threshold, which in clinical situations is the unknown variable, it was difficult to use the formulation of the equation suggested by Brambilla et al (21) in its original form. Erdi et al (28) obtained the approximate lesion size from the CT and T/B ratio from PET to obtain the appropriate optimum threshold which was applied to the PET images to obtain lesion activity and a final estimate of the lesion volume.…”