Patient data
Dynamic FDG PET scans from 10 non-small cell lung cancer [NSCLC] (stages IIIB to IV) patients [12] and 8 gastrointestinal [GI] (colorectal carcinoma) cancer patients [13] were included retrospectively. All scans had been acquired prior to therapy. All patients had given written informed consent, and both studies had been approved by the Medical Ethics Review Committee of the VU University Medical Center.
For NSCLC patients (three females, seven males; weight 76 ± 10 kg, range 56 to 94 kg), blood glucose levels were within the normal range (mean 5.5 ± 0.6 mmol·L-1, range 4.4 to 7.0 mmol·L-1). The same was true for blood glucose levels (mean 5.6 ± 0.8 mmol·L-1, range 3.9 to 7.0 mmol·L-1) of patients with advanced GI malignancies (one female, seven males; weight 85 ± 15 kg, range 60 to 110 kg).
PET scanning protocol
All patients fasted for at least 6 h before scanning. Patients were prepared in accordance with recently published guidelines for quantitative PET studies [14]. They were scanned in a supine position and received an intravenous catheter for tracer administration. During dynamic scanning, blood samples for determining plasma glucose levels were collected at fixed times (i.e., at 35, 45, 55 min post injection). All dynamic scans were performed using an ECAT EXACT HR+ scanner (Siemens/CTI, Knoxville, TN, USA) [15], having a 15.5-cm axial field of view. Each scan session started with a 10-min transmission scan using three retractable rotating 68Ge line sources. After completion of the transmission scan, a bolus of FDG was administrated intravenously (388 ± 71 and 459 ± 97 MBq for NSCLC and GI cancer, respectively), at the same time starting a dynamic emission scan in a 2-D acquisition mode. Each dynamic scan consisted of 40 frames with the following lengths, 1 × 30, 6 × 5, 6 × 10, 3 × 20, 5 × 30, 5 × 60, 8 × 150, 6 × 300 s. In addition, a static scan was created by summing the sinograms of the last three frames (i.e., 45 to 60 min post injection).
All data were normalized and corrected for attenuation, random coincidences, scatter radiation, dead time, and decay. Reconstructions were performed using normalization and attenuation-weighted ordered subsets expectation maximization [OSEM] with 2 iterations and 16 subsets, followed by post-smoothing using a 0.5 Hanning filter. This resulted in an image resolution of approximately 6.5 mm full width at half maximum. An image matrix size of 256 × 256 × 63 was used, corresponding to a pixel size of 2.57 × 2.57 × 2.43 mm3.
After reconstruction, the summed image (45 to 60 min post injection) was used to generate a SUV image by normalizing local tissue concentrations to injected dose and body weight. In addition, Patlak analysis, a kinetic linearized model [16] for irreversible tracer uptake, was applied to the interval 10 to 60 min post injection to generate an image of net influx rate [K
i] of FDG, which is proportional to the metabolic rate of glucose. Image-derived input functions [IDIF] were used as plasma input curves and obtained as described by Cheebsumon et al. [17]. In short, 3-D volumes of interest [VOI] were drawn manually in three vascular structures (i.e., the left ventricle, aortic arch, and ascending aorta) using an early frame that highlights the blood pool [18]. These VOI were then projected onto all frames yielding arterial whole blood time activity curves (i.e., IDIF). The average input curves from VOI defined in the three vascular structures were used as an input function during Patlak analysis.
Data analysis
For the 10 NSCLC patients, VOI were defined for 54 lesions that were all located in the lung. For the 8 GI cancer patients, VOI were defined for 37 lesions that were located in the liver (n = 23), lung (n = 12), or colon (n = 2). All lesions that could be identified by an expert physician were included in this study. Metabolic tumor volumes were obtained using the following five different types of (semi-)automatic VOI methods:
Fixed threshold of 50% and 70% (VOI50
, VOI70
). In this method, a fixed threshold (i.e., 50% or 70%) of the maximum voxel value within a tumor is used to delineate the tumor [19].
Adaptive threshold of 41%, 50%, and 70% (VOIA41
, VOIA50
, VOIA70
). This 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 the tumor and local background [19].
Contrast-oriented method (VOISchaefer
). This method uses the average of SUV within a 70% threshold of SUVmax isocontour (meanSUV70%) and background activity for various sphere sizes. Regression coefficients are calculated, which represent the relationship between the optimal threshold and image contrast for various sphere sizes [3]. This threshold equation is given by:
where A and B are fitted using phantom studies [3]. When applied to Patlak images, K
i rather than SUV is used. In general, different values are applied for sphere diameters smaller and larger than 3 cm in diameter. In the present study, this method was recalibrated, i.e., specific A and B values for the image characteristics used were determined.
Background-subtracted relative-threshold level [RTL] method (VOIRTL
). This method is an iterative method that is based on a convolution of the point-spread function, which takes into account differences between various sphere sizes and the scanner resolution [4].
Gradient-based watershed segmentation method. This method uses two steps before calculating the VOI [2]. First, a gradient image is calculated on which a 'seed' is placed in the tumor (tumor basin) and another in the background (background basin). Next, a watershed [WT] algorithm is used to grow both seeds in the gradient basins, thereby creating boundaries on the gradient edges. In the present study, two different types of gradient basins were used. In the first approach [GradWT1], all voxels on the edge between the tumor and background are assigned to the tumor [8, 10]. In the second approach [GradWT2], an upsampled image is used to ensure less effects of sampling. In addition, a voxel on the edge between the tumor and background is allocated to either the tumor or background based on the smallest difference with that voxel value.
To reduce sensitivity to noise, for all methods, the maximum voxel value was obtained using a cross-shaped pattern. This method searches for the region with the (local) average maximum intensity based on the average of seven neighboring voxels, which was then used as a maximum or 'peak' value. The tumor-to-background ratio was calculated by dividing this maximum value by the background value surrounding the tumor.
Statistical analyses
Both metabolic volumes and differences in measured volumes derived from two image types are reported. The percentage volume difference was defined as . Note that this value can be negative, indicating an underestimation of the SUV-derived metabolic volume compared with the Patlak-derived volume. In addition, for each delineation method, mean, median, first quartile, third quartile, minimum, and maximum values, including statistical outliers, are reported in box plots. Moreover, visual outliers were identified as VOIs that showed unrealistically large or small volumes when compared visually with the tumor. These outliers were not included in the statistical analysis when calculating p values. A two-tailed Wilcoxon signed-rank test was used to indicate statistically significant differences between measured volumes derived from SUV and Patlak images, where p values less than 0.05 were considered to be significant.