Page 33 - Towards personalized therapy for metastatic prostate cancer: technical validation of [18F]fluoromethylcholine

Dual-phase [18F]FCH PET/CT in prostate cancer
For lesion delineation, we used the adaptive threshold of 50% of maximum voxel value
within tumor, the 3D volume of interest A50 (VOIA50). This method is similar to the
fixed threshold method, except that it adapts the threshold relative to the local average
background, thereby correcting for the contrast between tumor and local background.
For example, the A50 contour value corresponds to a value at 50% of the sum of the
maximum voxel value and the local background value. The latter value is derived from 2 ‘background’ voxels that are identified as those voxels located on a single voxel thick
shell at 2.5 cm from the edge of a 70% of maximum pixel value isocontour, excluding all voxels with an SUV larger than 2.5.
SUV’s were normalized for body weight. For data-analysis, we used early (2 min p.i.) and late (30 min p.i.) SUVmax and SUVmean, as well as their absolute [SUV late - SUV early] and relative differences [(SUV late - SUV early)/ SUV early].
Furthermore, based on literature [8], we considered two possible time-trends of [18F] FCH uptake by comparing the SUV early and the SUV late: type A pattern (decreasing over time) if the SUV early > the SUV late and type B pattern (stable/ increasing over time) if the SUV late was equal or exceeded the SUV early. Equality of SUV’s was decided using the 2nd decimal. In the present context, accuracy measures relate to the ability of time-trends (2 versus 30 min p.i.) of tracer uptake to discriminate malignant and benign lymph nodes with enhanced [18F]FCH uptake.
Statistical Analysis
The two sample Mann-Whitney tests were performed to determine a shift in the median values for benign and malignant tumors. Linear mixed-effects models were constructed to determine the relation between SUVmax and SUVmean. Included in this model were tumor status (benign versus malignant) and post injection scan time (2 min versus 30 min) as fixed effects, as well as random slopes per patient and per lesion (nested within patient). The residuals were assumed to be exponential related to SUVmean, and pairwise interactions between SUVmean and both tumor status and scan time were also tested. An identical linear regression model ignoring repeated measures and heteroscedasticity was used to provide an R2. Receiver Operating Characteristic (ROC) curves were constructed to determine the thresholds maximizing specificity and sensitivity. The binomial distribution is used to determine 95% confidence intervals of the sensitivity and specificity estimates for the optimal thresholds. To ensure that these results are not influenced by within-patient correlations in uptake, a Monty
31