Influence of different contributions of scatter and attenuation on the threshold values in contrast-based algorithms for volume segmentation

Article Outline

• Abstract
• Introduction
• Materials and methods
  • Scanner and phantoms
  • Source preparation and acquisition protocol
  • Image reconstruction and data analysis
  • Statistical analysis
• Results
  • Cross section A ≤ 133 mm2
  • Cross section A > 133 mm2
• Discussion
• Conclusion
• Acknowledgements
• References
• Copyright

Abstract 

The aim of this work is to evaluate the role of different amount of attenuation and scatter on FDG-PET image volume segmentation using a contrast-oriented method based on the target-to-background (TB) ratio and target dimensions. A phantom study was designed employing 3 phantom sets, which provided a clinical range of attenuation and scatter conditions, equipped with 6 spheres of different volumes (0.5–26.5 ml). The phantoms were: (1) the Hoffman 3-dimensional brain phantom, (2) a modified International Electro technical Commission (IEC) phantom with an annular ring of water bags of 3 cm thickness fit over the IEC phantom, and (3) a modified IEC phantom with an annular ring of water bags of 9 cm. The phantoms cavities were filled with a solution of FDG at 5.4 kBq/ml activity concentration, and the spheres with activity concentration ratios of about 16, 8, and 4 times the background activity concentration. Images were acquired with a Biograph 16 HI-REZ PET/CT scanner. Thresholds (TS) were determined as a percentage of the maximum intensity in the cross section area of the spheres. To reduce statistical fluctuations a nominal maximum value is calculated as the mean from all voxel >95%. To find the TS value that yielded an area A best matching the true value, the cross section were auto-contoured in the attenuation corrected slices varying TS in step of 1%, until the area so determined differed by less than 10 mm² versus its known physical value. Multiple regression methods were used to derive an adaptive thresholding algorithm and to test its dependence on different conditions of attenuation and scatter.

The errors of scatter and attenuation correction increased with increasing amount of attenuation and scatter in the phantoms. Despite these increasing inaccuracies, PET threshold segmentation algorithms resulted not influenced by the different condition of attenuation and scatter. The test of the hypothesis of coincident regression lines for the three phantoms used provided no statistical basis for believing that the three lines are not coincident.

Calibration curves needed to implement contouring algorithms based on adaptive TS segmentation of PET volumes can be devised in different conditions of attenuation and scatter. This opens the possibility of defining a unified contrast-based method for target delineation in different anatomical districts.

Keywords: FDG-PET/CT, Radiation treatment planning, Functional imaging, Target volume definition

Back to Article Outline

Introduction 

Although increasingly used in the clinical practice, the determination of tumour volume and shape from ¹⁸F-flourodeoxyglucose (FDG) positron emission tomography (PET) in the course of radiotherapy planning remains a challenging task. Various approaches were reported in the literature to accurately contour FDG based gross target volume (GTV) [1]. Visual contouring by an experienced observer remains observer dependent [2]. Absolute thresholds such as a standardized uptake value (SUV) of a fixed value, e.g. 2.5, surrounding the lesion often fails when the physiological background activity lies above the fixed threshold [3]. Also the use of mean target SUV [4], or the use of a fixed percentage threshold, e.g. 40–50%, relative to maximum FDG accumulation of the lesions [5] are not suitable for GTV contouring [6]. Methods for segmentation based on contrast-oriented contouring algorithms [3], [7], [8], [9] have been developed independently by many groups and validated in patient data both in head and neck [10] and in lung cancer [11] with satisfactory results. Methods for segmentation of non uniform tracer concentration not based on thresholding have been recently proposed [12], [13]. While referring to these promising methods it should be pointed out that until they are further developed and validated, contrast-oriented segmentation methods are and will be used in most clinics and therefore need to be accurately characterized.

The use of contrast-based segmentation methods is driven by the low quality of PET images in terms of resolution and statistical noise. These methods rely on a scanner-specific calibration curve, which depends on object properties such as target-to-background (TB) ratio [5], [7], [14] and target size [14], [15], [16], and imaging parameters such as reconstruction algorithm and smoothing filter [7], [15]. We recently demonstrated that, among acquisition parameters, emission scan duration and background activity concentration, both related to total number of counts and to the level of image noise, did not result as significant predictors in threshold determination in the ranges explored [14]. The influence of different conditions of attenuation and scatter on contrast-based algorithms, used to define the boundaries of FDG uptake has not been fully explored in literature. This is of clinical relevance since, if different conditions of attenuation and scatter would play a role, site-specific calibration curves should be devised at least for a coarse subdivision of whole-body imaging such as, for instance, head and neck, thorax and abdomen. On the other hand, the finding that PET image segmentation is independent on scatter and attenuation would suggest that contrast-based contouring algorithms need not to be regarded as site specific and may be applied irrespective of the phantom used in their derivation.

In hybrid PET/CT scanner, attenuation correction is performed using CT data, scaled to 511 keV [17]. Several approaches to the problem of scatter correction have been proposed and are currently used including analytical corrections [18], Monte Carlo simulations [19], background subtraction or tail fitting methods [20]. A standardized way to quantify the accuracy of correction for attenuation and scatter has been provided in the NEMA NU2-2001 (NEMA01) standard [21] as the relative error in percentage units of the intensity within a region of the lung insert of the International Electro Technical Commission (IEC) body phantom set compared to the background intensity: values of 32% and 13% have been reported by Erdi et al. [22] using a CPS Accel PET/CT scanner and attenuation-weighted ordered-subset expectation maximization (OSEM) iterative reconstruction with 2 iterations and 8 subsets and 4 iterations and 16 subsets, respectively. Similar values (34% and 17%) were obtained by our group using a Biograph HI-REZ PET/CT scanner [23]. Bettinardi et al. [24] reported a value of 16% using a Discovery ST PET/CT scanner in 3D acquisition mode and OSEM (3 iterations, 32 subsets) reconstruction. In a recent inter-laboratory comparison study of image quality of PET scanner, Bergmann et al. [25] reported relative errors for the attenuation and scatter correction, for 8 dedicated PET systems for which attenuation correction was routinely used, in a range of 26–49%. Moreover, they demonstrated a significant negative correlation of this relative error with a contrast quality index, defined as the sum of all contrast values measured on the 6 IEC phantom spheres.

Our aim is to show that, also with simplified phantom geometries as those selected in this study, the accuracy of correction decreases with increasing amount of attenuation and scatter in the phantom being imaged. To this purpose an experiment was designed in which the accuracy of correction for attenuation and scatter was assessed in three phantom sets representing different conditions of attenuation and scatter, assessed by conventional measures of noise equivalent count rate (NECR) and scatter fraction characteristics. We previously provided evidence that the modification of the IEC phantom used in the experiment generates counting rates that are relevant to predict what would happen in human imaging [26]. Based on these results, an adaptive thresholding algorithm based on contrast and lesion size was developed. The impact of different conditions of attenuation and scatter on the algorithm was assessed for a wide range of TB ratios, target sizes and image reconstruction parameters using the same phantom sets.

Back to Article Outline

Materials and methods 

Scanner and phantoms 

Images were acquired with the Biograph 16 HI-REZ PET/CT scanner (Siemens Medical Solutions), equipped with Pico-3D digital electronics. A 16.2 cm axial field of view is covered by 81 image planes with slice thickness of 2 mm for each bed. The scanner transverse spatial resolution and axial resolution are 4.6 and 5.1 mm FWHM a 1 cm radial position. Both axial and transaxial FWHM values degraded by about 0.8 mm when moving from 1 to 10 cm away from the central axis of the scanner [23]. The CT portion of the scanner is the Somatom Sensation sixteen-slice CT and is used both for attenuation correction of PET data and for anatomic localization of FDG uptake in fused PET/CT images.

Measurements were performed on 3 phantoms sets (Data Spectrum Corporation). The smallest (Phantom 1) is the Hoffman 3-D brain phantom, used without brain inserts and approximately equal in size to the head of an adult. To assess the accuracy of attenuation and scatter correction in the Hoffman phantom, a 5 cm diameter plastic insert packed with water and foam beds and with a density approximately equal to the average value in lung tissue (0.30 g/ml) was placed in the center of the phantom. Hounsfield units of the lung inserts averaged −602 ± 92 HU. The next (Phantom 2) is a modification of the International Electro technical Commission (IEC) body phantom set: the IEC phantom alone tends to over-estimate true and random count rates and to under-estimate scatter fraction common to clinical patient scanning [21], and so additional attenuation and scatter material (an annular ring of 3 cm water bags) were added to better approximate typical clinical count rates. The largest one (Phantom 3) was obtained by fitting an annular ring of water bags of 9 cm thickness over the IEC phantom. In each phantom 6 fillable polymethilmetacrilate (PMMA) spheres with internal diameter of: 10, 13, 17, 22, 28 and 37 mm and wall thickness of 1 mm were inserted. To simulate the presence of activity external to the field of view, a NEMA01 scatter phantom set was used including a solid polyethylene cylinder with a 0.96 g/cm³ density, an outside diameter of 20.3 cm, an overall length of 70 cm and a fillable 80 cm long plastic tube with a 3.2 mm ID inserted in a hole drilled parallel to the central axis of the cylinder at a 4.5 cm radial distance. The scatter phantom was positioned at the end of the phantoms used in every acquisition. The plastic tube was filled so to have an equivalent activity concentration in the whole scatter phantom as the one of the main chamber of the phantom imaged, as requested by the NEMA01 standard and to approximate an average condition that can be encountered clinically [26]. In the scatter fraction and the NECR test, a fillable plastic tubes of 3.2 mm internal ID was inserted in the main chamber of the phantoms, parallel to the central axis of the cylinders at 4.5 cm radial distance, so that the line sources were at the position nearest the patient table. The experimental setup of Phantom 2 acquisitions is shown in Fig. 1. The characteristics of the phantoms are given in Table 1.

• 
  • View Large Image
  • Download to PowerPoint

• Figure 1 

  CT scout view of Phantom 2 and NEMA scatter phantom in the experimental setup (A), CT axial view of Phantom 2 (B).

Table 1. Characteristics of the phantoms used.

Source preparation and acquisition protocol 

A NEMA01 approach was used to determine the scatter fraction and the NECR of the phantoms used. The phantoms were placed centered both in the transverse and axial field of view. The main chamber and all the spheres of the phantoms were filled with non radioactive water. The line sources inside the main chamber of the phantoms were filled with activity, so to provide a starting activity concentration in whole phantom ranging from 36 to 73 kBq/ml. Forty six temporal image frames were acquired: the first ten with 10-min duration, each followed by a 5-min gap; the second ten frames had 15-min duration, and the last 26 frames had 20-min durations, each of the last 26 frames followed by 10-min gaps. Data were acquired over 18 h and each acquisition contained a minimum of 2.5 million prompt counts. Separate prompt and delayed sinograms were acquired using a standard delayed coincidence window technique.

In a second test, the background for each phantom was filled with 5.4 kBq/ml activity concentration. The source-to-background ratios, as determined by the dose calibrator, were set to 4, 8 and 16 for each phantom in three different acquisition sessions. Overall, 9 (3 phantoms × 3 source-to-background ratios) statistically independent fully 3D coincidence sinograms were acquired. TB ratios were also determined in the reconstructed image as the maximum pixel intensity in a region-of-interest (ROI) encircling the cross sectional area of the target, divided by the average pixel intensity of ROIs surrounding the sphere. ROIs analysis was performed, as previously described [26], by means of an automatic routine, developed using IDL 6.1 (Research System, Inc.) to avoid the influence of the operator in ROIs dimensioning and to minimize the influence of the operator in the ROIs positioning. Briefly, a pattern of six ROIs of fixed dimensions (diameters equal to the physical ID of the spheres) and fixed relative distances is presented to the operator who can only rotate and translate the pattern to establish its correct position over the hot spheres in the slice. The ROIs analysis tool permits movement of the ROIs pattern in increments of less than 1 mm. The operator is also requested to position a pattern of twelve 37 mm background ROIs at a distance of 15 mm from the edge of the phantom but no closer than 15 mm to any sphere. The positioning and dimensioning of the smaller ROIs (10, 13, 17, 22, and 28 mm) on background were done automatically from the placement of the original 12 background ROIs. The same pattern of 12 background ROIs was automatically positioned at a distance of ±1 and ±2 cm from the central slice for a total of 60 background ROIs, as prescribed by NEMA recommendations. The same analysis is repeated for the four micro-hollow spheres by choosing a different central slice.

The total activities present in the phantom setup were in the range of 130–170 MBq. The emission scan duration was set to 5 min/bed according to clinical acquisition protocols used in our institution for radiotherapy planning.

Image reconstruction and data analysis 

3D sinograms were rebinned by using the single slice-rebinning algorithm. Data analysis followed the procedure suggested by Watson et al. [27]. System scatter fraction is reported at low activity levels. The NECR curve was generated for k = 2 (noisy random correction), as is commonly done when randoms are determined from a measured delayed coincidence window.

Accuracy of attenuation and scatter correction is calculated for each slice as the relative error in percentage units of the intensity within a region of the lung insert compared to the background intensity, following the procedure detailed in the NEMA01 standard. The relative error for slice i is given by:

With c_lung,i the mean intensity in a region within the lung insert and c_bkgd,i the mean of the intensities of 60 background regions drawn for the image quality analysis of the NEMA01 standard. The mean relative error is calculated as the mean of the individual relative errors of individual slices.

Data were corrected for random coincidences, normalization, dead-time losses, scatter and attenuation. Image reconstruction was performed using the attenuation-weighted-OSEM iterative reconstruction (two iterations, eight subsets, 4 mm Gaussian filter). The data were reconstructed over a 256 × 256 matrix providing a voxel size of 2.62 × 2.62 × 2 mm.

Thresholds (TS) were determined as a percentage of the maximum intensity in the cross section area of the spheres. To reduce statistical fluctuations a nominal maximum value is calculated as the mean from all voxel >95%. Target cross sections of area A were selected precisely in the middle of the spheres, which constitutes the largest cross section of the sphere, by using the inherently co-registered CT scan. The values of TS were entirely based on the apparent activity concentration in the images and not on the known activities actually placed in the spheres. To find the TS value that yielded an area A best matching the true value, the cross section were auto-contoured in the attenuation corrected slices varying TS in step of 1%, until the area so determined differed by less than 10 mm² versus its known physical value.

To further investigate whether there is an influence of the number of iterations and smoothing kernel on the results, an additional series of reconstructions was performed realizing a comparably over converged and smoothed iterative reconstruction (8 iterations, 8 subsets, 8 mm Gaussian filter).

Statistical analysis 

Differences in the average error in scatter and attenuation corrections among the three phantom sets were assessed using analysis of variance methods. An F test was made of the null hypothesis (H₀) that there are no differences among the average errors. Post-hoc comparisons among the means of the three phantom sets were performed with the Tukey's test.

The relationship between the best TS (Y) that provides the most accurate cross sectional area of the spheres and the variables X₁ (defined as 1-1/TB) and X₂ (defined as target cross section A), both linearly related to Y [14] was established using stepwise multiple linear regression methods [28], using the starting model:

Where B₀, B₁ and B₂ are the regression coefficients that need to be estimated and E is the error term. Goodness of fit was expressed using the adjusted coefficient of determination (R²), which is the proportion of variability in a data set that is accounted for by the statistical model and provides a measure of how well future outcomes are likely to be predicted by the model. The weight of different X-variables in explaining Y was quantified by means of standard regression coefficients β_i, which can be used as a measure of relative importance, with the Xs ranked in order of the absolute size of these coefficients.

Suppose now that we wish to compare the separate multiple regressions of TS on A and (1 − 1/TB) for the three phantoms. For each case in each phantom group, we observe values of the variable Y = TS, X₁ = A and X₂ = (1 − 1/TB); further, we shall suppose that there are n_i cases in the ith Phantom group, i = 1,2,3. We begin by defining two dummy variables Z₁ and Z₂ as follows: Z₁ = 1 if Phantom 2 case; 0 otherwise; Z₂ = 1 if Phantom 3 case; 0 otherwise.

The complete model to be used is then given as follows:

For each particular Phantom, model (3) specializes as follow:

The following hypotheses concerning the parameters in model (3) are of interest:

Are all the three regression equations coincident (i.e, test H₀: B₃ = B₄ = B₅ = B₆ = B₇ = B₈ = 0)? When H₀ is true, all three phantom models reduce to the form:

The test statistic is the multiple partial F:

where:

residual SS = residual sum of squares; MS = mean square; ν = number of independent linear parametric functions specified to be 0 under H₀. H₀ must be tested using Analysis of Variance tables with ν = 6 and n1 + n2 + n3 − ñ degrees of freedom, being ñ the number of parameters in the full model.

Statistical analysis was performed using the software Statistica 6.0 (Statsoft Inc, USA).

Back to Article Outline

Results 

Scatter fractions are reported in Table 2 for the Hoffman and modified IEC phantoms. Peak NECR for the combination of Hoffman and modified IECs plus scatter phantom are reported in the same table, while corresponding NECR curves are plotted in Fig. 2. The scatter fractions measured at low activity levels, were 27.4%, 40.4% and 48.7% for the Hoffman, and modified IECs phantoms. The peak NECR of 136.6 kcps was almost reached in the Hoffman phantom at 70.9 kBq/ml. As expected, due to increasing phantom diameter, peak NECR are lower and are reached at a lower activity concentration for the modified IEC with 3 cm additional thickness (61.2 kcps at 31.8 kBq/ml) and modified IEC with 9 cm additional thickness (∼30.8 kcps at 25.2 kBq/ml).

Table 2. Physical properties of the phantoms.

aPeak NECR was not reached due to insufficient starting activity.

• 
  • View Large Image
  • Download to PowerPoint

• Figure 2 

  NECR (k = 2) curves for the Hoffman and modified IEC plus scatter phantoms.

The error of correction for attenuation and scatter increases with increasing amount of attenuation and scatter in the phantom being imaged, as shown in Table 2. Overall, this increase is statistically significant (F = 406.8 p < 10⁻⁶). Individual significant differences (p = <10⁻⁴) were found between Phantom 1 (ΔC_lung = 23.7 ± 1.6%) and Phantom 2 (ΔC_lung = 37.6 ± 3.4%) and between Phantom 2 and Phantom 3 (ΔC_lung = 40.8 ± 5.1%) (p = <10⁻⁴).

The partial-volume effects and spill over influence the measured source activity concentration in the sphere: the measured TB ratio obtained from PET images differed from prepared source-to-background ratio as determined by the dose calibrator. These TB ratios ranged from 22.1 down to 1.6 and were within the range observed in patients with lung or head and neck tumours.

The TS versus cross sectional area relationship was splitted and fitted into different functional forms as already performed in previous studies [14], [16]. Partial-volume effects significantly reduce the contrast recovery for structures less than three-times the reconstructed image resolution [29] which in our scanner is about of 4.5 mm. Thus the choice of sphere A = 133 mm² (or, equivalently, a sphere internal diameter of 13 mm) as a separator of the data was dictated by the resolution characteristics of our scanner.

Cross section A ≤ 133 mm² 

The regression equation that best summarizes the results obtained in a multiple regression model with TS as predicted variable and (1 − 1/TB) and sphere A as predictor variables may be written as:

The adjusted coefficient of determination is R² = 0.80. Both sphere A and (1 − 1/TB) resulted as statistically significant predictors of TS. TS diminishes with increasing sphere A and with increasing TB ratios. The functional dependence of TS on (1 − 1/TB) is higher (β_(1−1/TB) = −0.67) than the influence of A (β_A = −0.55) for the small volumes corresponding to sphere A ≤ 133 mm². The test of the hypothesis of coincident regression lines for the three phantoms provided an F_6,13 = 0.09 (P = 0.996). This F statistic is extremely small (P is very large); so we do not reject H₀ and therefore have no statistical basis for believing that the three lines are not coincident. The details of the test are reported in Table 3.

Table 3. ANOVA table. Test of H₀ = coincident regression lines for the three phantoms. A≤133 mm².

.

Cross section A > 133 mm² 

The regression equation that best summarizes the results obtained in a multiple regression model with TS as predicted variable and TB ratio and sphere A as predictor variables may be written as:

The adjusted coefficient of determination is R² = 0.73. Only (1 − 1/TB) resulted as a statistically significant predictor of TS, while sphere A is no longer retained into the model. Again TS diminishes with increasing TB ratios. The test of the hypothesis of coincident regression lines for the three phantoms provided an F_4,30 = 1.03 (P = 0.408). This F statistic is small (P is large); so we do not reject H₀ and therefore have no statistical basis for believing that the three lines are not coincident. The details of the test are reported in Table 4.

Table 4. ANOVA table. Test of H₀ = coincident regression lines for the three phantoms. A> 133 mm².

.

Similar results were found when multiple regression analysis was lead over the additional series of OSEM reconstruction with 8 iteration, 8 subsets and 8 mm Gaussian filter.

Back to Article Outline

Discussion 

Attenuation and scatter corrections are routinely performed during image reconstruction in PET/CT scanner. Many authors [22], [23], [24], [25] have previously demonstrated that the errors in attenuation and scatter corrections are substantial, roughly ranging from 15% to 50%, and with a significant negative correlation with a contrast quality index [25]. Notwithstanding the simplified phantom geometries used in this study, we also provided evidence in Table 2 that the error in scatter and attenuation correction increases with increasing amount of attenuation and scatter in the phantom being imaged. The increase in lung error is clearly significant (∼14%) when passing from a head and neck geometry, simulated through a Hoffman phantom, to a body geometry (modified IEC phantoms). Thus, an analysis of the role of attenuation and scatter on threshold determination is warranted. To this purpose, an experiment was designed in which threshold calibration curves were derived in three phantom sets representing different conditions of attenuation and scatter.

The first step to study this dependence in conditions resembling the ones that can be encountered in clinical studies is to provide evidence that the simulated conditions of attenuation and scatter adequately cover the range that can be observed in patient imaging. Watson et al. [27] using a very similar scanner (only crystal dimensions are different but this impacts on resolution not on scatter properties) demonstrated that mean scatter fraction increased only moderately over the entire range of patient weight examined (35–135 kg) with a mean value of 44% corresponding to 80-kg patient. While the Hoffman phantom is intended to simulate the head, the modified IEC phantom sets adequately cover the entire range of scatter fractions that can be encountered in clinical whole-body imaging. We had previously demonstrated that the second model adopted, a modified IEC phantom plus the NEMA01 scatter phantom, approximate typical clinical count rates [26]. Altogether, these findings provide evidence that the models used in the experiment generates counting rates that are relevant to predict what would happen in human imaging and that we are spanning the entire range of attenuation and scatter conditions that can be encountered in clinical imaging.

Another relevant point that deserves further consideration is the choice of a cross section for the analysis instead of a volume approach. As shown by Drever et al. [16], no single threshold value will simultaneously yield both a correct determination of total target volume and also individual cross section for objects of variable cross sectional shape. Only a slice specific threshold level can come closest to correctly reproducing both correct cross section and the total volume of a structure. Although spherical objects were used in the present study to derive the coefficients of the algorithm (and in this configuration a cross section and a volume approach are likely to come to the same conclusion) the question of how properly define target volumes is crucial to the efficacious quantitative use of the functional information provided by PET in radiotherapy planning, and for future validation of the algorithm both in phantom and in patients. The practical utility of an optimal single threshold derived from a volumetric approach is questionable at best as it would apply only to spherical volumes. It must be underlined that both approaches provided a volume determination. The only difference is that in the slice specific approach the volume is determined as the sum of different slabs in which segmentation is based upon the slice specific contrast between target and surrounding background.

As a second step, a reasonable model must be defined before running multiple regression analysis. In this study, the focus was on the role played on TS determination by different conditions of attenuation and scatter, simulated through the use of different phantom sets. The different phantom models represent a categorical predictor whose influence on the dependent variable TS must be ascertained. The methods of regression analysis can be generalized to treat categorical predictors through the use of dummy variables. In this paper we focused on an important application of dummy variables: comparing several regression equations by use of a single multiple regression model. From a previous study [14], we know that among a restricted set of independent variables related to object characteristics (target dimensions and TB) or acquisition parameters (background activity concentration and emission scan duration) only the former explains a significant amount of TS variance. In the framework of multiple regression analysis, cross section area A and TB were assumed as continuous predictor variables, while phantoms, or equivalently different conditions of attenuation and scatter, were represented using two dummy categorical variables. The same analysis was repeated using a comparably over converged and smoothed iterative reconstruction to ascertain whether there is an impact of the number of iterations and smoothing kernel on the results. No other basic independent variables were controlled and inserted into the analysis.

With these premises, the major finding of this work is the demonstration that different conditions of attenuation and scatter do not influence TS in the wide range explored of TB ratios, target sizes and reconstruction parameters. Adaptive thresholding algorithms rely on calibration curves, which depend on parameters such as object characteristics (e.g. TB ratio, target size) and reconstruction parameters (e.g. algorithms, smoothing filter). Others parameters have been advocated as potential predictors of TS. Among them, the phantom used in the derivation of the calibration curve [11] or equivalently the effects of attenuation and scatter [15] were demonstrated in the present work not to be related with TS. This, together with the previously demonstrated independence of TS on acquisition parameters such as emission scan duration or background activity concentration [11], [14], adds to the generalizability of this methods: distinct calibration curves need not to be specifically devised for each anatomical site representing different conditions of attenuation and scatter and, on the other hand, curves generated in one PET center for well defined conditions can be exported to other centers using different working conditions in terms of emission scan duration, administered activities or delay between injection and acquisition, provided that both are equipped with the same scanner and use the same reconstruction parameters.

A study limitation must be acknowledged: the study was not planned to accurately define an adaptive thresholding algorithm. To this purpose a significantly greater number of observations should have been collected mainly to better span the entire range of TB and target dimensions that can be found clinically and possibly improve the coefficient of determination that in our experiment only accounted from 73% to 80% of the TS variance, depending on the lesion size.

Back to Article Outline

Conclusion 

Different conditions of attenuation and scatter do not play a significant role in explaining the variance of TS the wide range explored of TB ratios, target sizes and reconstruction parameters. This, in turn ensures that calibration curves needed to implement contrast-based contouring algorithms of PET volumes can be devised in different conditions of attenuation and scatter. This opens the possibility of defining a unified contrast-based method for target delineation in different anatomical districts.

Back to Article Outline

Acknowledgements 

This work was supported by a grant from the Piedmont Region, Italy in the frame of “Ricerca Sanitaria Finalizzata 2008”.

Back to Article Outline

References 

1. Nestle U, Weber W, Hentschel M, Grosu AL. Biological imaging in radiation therapy: role of positron emission tomography. Phys Med Biol. 2009;54:R1–R25
   • View In Article
   • CrossRef
2. Riegel AC, Berson AM, Destian S, et al. Variability of gross tumor volume delineation in head-and-neck cancer using ct and PET/CT fusion. Int J Radiat Oncol Biol Phys. 2006;65:726–732
   • View In Article
   • Abstract
   • Full Text
   • Full-Text PDF (262 KB)
   • CrossRef
3. Nestle U, Kremp S, Schaefer-Schuler A, et al. Comparison of different methods for delineation of ¹⁸F-FDG PET-positive tissue for target volume definition in radiotherapy of patients with non-Small cell lung cancer. J Nucl Med. 2005;46:1342–1348
   • View In Article
   • MEDLINE
4. Black QC, Grills IS, Kestin LL, et al. Defining a radiotherapy target with positron emission tomography. Int J Radiat Oncol Biol Phys. 2004;60:1272–1282
   • View In Article
   • Abstract
   • Full Text
   • Full-Text PDF (1048 KB)
   • CrossRef
5. Erdi YE, Mawlawi O, Larson SM, et al. Segmentation of lung lesion volume by adaptive positron emission tomography image thresholding. Cancer. 1997;80:2505–2509
   • View In Article
   • CrossRef
6. Bayne M, MacManus M, Hicks R, Leong T, Peters L, Ball D. Can a mathematical formula help define a radiation target volume using positron emission tomography? In regard to Black et al. 2004. Int J Radiat Oncol Biol Phys. 2005;62:299–300
   • View In Article
   • Full Text
   • Full-Text PDF (52 KB)
   • CrossRef
7. Daisne JF, Sibomana M, Bol A, Doumont T, Lonneux M, Gregoire V. Tri-dimensional automatic segmentation of PET volumes based on measured source-to-background ratios, influence of reconstruction algorithms. Radiother Oncol. 2003;69:247–250
   • View In Article
   • Abstract
   • Full Text
   • Full-Text PDF (108 KB)
   • CrossRef
8. Davies JB, Reiner B, Huser M, Burger C, Szekely G, Ciernik IF. Assessment of 18F PET signals for automatic target volume definition in radiotherapy treatment planning. Radiother Oncol. 2006;80:43–50
   • View In Article
   • Abstract
   • Full Text
   • Full-Text PDF (852 KB)
   • CrossRef
9. Jentzen W, Freudenberg L, Einsing EG, Heinze M, Brandau W, Bockisch A. Segmentation of PET volumes by iterative image thresholding. J Nucl Med. 2007;48:108–114
   • View In Article
   • MEDLINE
10. Daisne JF, Duprez T, Weynand B, et al. Tumor volume in pharyngolaryngeal squamous cell carcinoma: comparison at CT, MR imaging, and FDG PET and validation with surgical specimen. Radiology. 2004;233:93–100
    • View In Article
    • MEDLINE
    • CrossRef
11. Schaefer Kremp S, Hellwig D, Rube C, Kirsch CM, Nestle U. A contrast-oriented algorithm for FDG-PET-based delineation of tumour volumes for the radiotherapy of the lung cancer: derivation from phantom measurements and validation in patient data. Eur J Nucl Med Mol Imaging. 2008;35:1989–1999
    • View In Article
12. Geets X, Lee JA, Bol A, Lonneux M, Gregoire V. A gradient-base method for segmenting FDG-PET images: methodology and validation. Eur J Nucl Med Mol Imaging. 2007;34:1427–1438
    • View In Article
13. Hatt M, Lamare F, Boussion N, et al. Fuzzy hidden Markov chains segmentation for volume determination and quantitation in PET. Phys Med Biol. 2007;52:3467–3491
    • View In Article
    • CrossRef
14. Brambilla M, Matheoud R, Secco C, Loi G, Krengli M, Inglese E. Threshold segmentation for PET target volume delineation in radiation treatment planning: the role of target-to-background ratio and target size. Med Phys. 2008;35:1207–1213
    • View In Article
    • CrossRef
15. Ford EC, Kinahan PE, Hanlon L, et al. Tumor delineation using PET in head and neck cancers: threshold contouring and lesion volumes. Med Phys. 2006;33:4280–4288
    • View In Article
    • MEDLINE
    • CrossRef
16. Drever L, Roa W, McEwan A, Robinson D. Iterative threshold segmentation for PET target volume delineation. Med Phys. 2007;34:1253–1265
    • View In Article
    • MEDLINE
    • CrossRef
17. Kinahan PE, Townsend DW, Beyer T, Sashin D. Attenuation correction for a combined PET/CT scanner. Med Phys. 1998;25:2046–2053
    • View In Article
    • MEDLINE
    • CrossRef
18. Vandervoort E, Sossi V. An analytical scatter correction for singles-mode transmission data in PET. IEEE Trans Med Imaging. 2008;27:402–441
    • View In Article
    • CrossRef
19. Barret O, Carpenter TA, Clark JC, Ansorge RE, Fryer TD. Monte Carlo simulation and scatter correction of the GE advance PET scanner with SimSET and Geant4. Phys Med Biol. 2005;50:4823–4840
    • View In Article
    • MEDLINE
    • CrossRef
20. Karp JS, Muehllehner G, Mankoff DA, et al. Continuous-slice PENN-PET: a positron tomograph with volume imaging capability. J Nucl Med. 1990;31:617–627
    • View In Article
    • MEDLINE
21. National Electrical Manufacturers Association . NEMA standards publication NU-2–2001: performance measurements of positron emission tomographs. Rosslyn (VA): NEMA; 2001;
    • View In Article
22. Erdi YE, Nehmeh SA, Mulnix T, Humm J, Watson CC. PET performance measurements for an LSO-based combined PET/TC scanner using the national electrical manufacturers association NU 2-2001 standard. J Nucl Med. 2004;45:813–821
    • View In Article
    • MEDLINE
23. Brambilla M, Secco C, Dominietto M, Matheoud R, Sacchetti G, Inglese E. Performance characteristics obtained for a new 3-dimensional lutetium oxyorthosilicate–based whole-body PET/CT scanner with the National Electrical Manufacturers Association NU 2-2001 Standard. J Nucl Med. 2005;46:2083–2091
    • View In Article
    • MEDLINE
24. Bettinardi V, Danna M, Savi A, et al. Performance evaluation of the new whole-body PET/CT scanner: discovery ST. Eur J Nucl Med Mol Imaging. 2004;31:867–881
    • View In Article
    • MEDLINE
25. Bergmann H, Dobrozemsky G, Minear G, Nicoletti R, Samal M. An inter-laboratory comparison study of image quality of PET scanners using the NEMA NU 2-2001 procedure for assessment of image quality. Phys Med Biol. 2005;50:2193–2207
    • View In Article
    • MEDLINE
    • CrossRef
26. Brambilla M, Matheoud R, Secco C, et al. Impact of target-to-background ratio, target size, emission scan duration and activity on physical figures of merit for a 3D LSO-based whole body PET/CT scanner. Med Phys. 2007;34:3854–3865
    • View In Article
    • CrossRef
27. Watson CC, Casey ME, Bendriem B, et al. Optimizing injected dose in clinical PET by accurately modelling the counting-rate response function specific to individual patient scan. J Nucl Med. 2005;46:1825–1834
    • View In Article
    • MEDLINE
28. Kleinbaum DG, Kupper LL, Muller KE. Applied regression analysis and other multivariable methods. Boston (Massachussets): PWS-Kent; 1987;
    • View In Article
29. Geworski L, Knoop BO, de Cabrejas ML, Knapp WH, Munz DL. Recovery correction for quantitation in emission tomography: a feasibility study. Eur J Nucl Med. 2000;27:161–169
    • View In Article
    • MEDLINE
    • CrossRef