Assessment of 18F PET signals for automatic target volume definition in radiotherapy treatment planning

Radiotherapy and Oncology 80 (2006) 43–50 www.thegreenjournal.com

Treatment planning

Assessment of 18F PET signals for automatic target volume definition in radiotherapy treatment planningq J. Bernard Davisa, Beatrice Reinera,*, Marius Husera,b, Cyrill Burgerc, Ga ´bor Sze ´kelyb, I. Frank Ciernika a

Radiation Oncology, University Hospital, Zurich, Switzerland, bComputer Vision Laboratory, Swiss Federal Institute of Technology, Zurich, Switzerland, cNuclear Medicine Department, University Hospital, Zurich, Switzerland

Abstract Introduction: Positron emission tomography (PET) alone or in combination with computer tomography (PET/CT) is increasingly used in target volume assessment. A standardized way of converting PET signals into target volumes is not available at present. Materials and methods: Assuming a uniform signal emission from a tumour and surrounding normal tissues, a modelbased method was developed to determine a relative threshold level (Threl) for gross tumour volume delineation. Two phantoms consisting of cylindrical and spherical sources of diameter ranging from 4.5 to 43 mm in a tank and 18F activities ranging from 0.001 to 0.15 MBq/ml for tank and sources, respectively, were used for PET/CT imaging. A Threl was calculated that best corresponded to the physical diameter of the cylindrical sources. Software (SW) was generated to automatically delineate volumes based on this threshold. The SW was validated for in vitro and in vivo PET signals. Results: The Threl best representing the source diameter was 41 ± 2.5% (95% confidence level) of the backgroundsubtracted signal. The mean deviation for sources of diameter P12.5 mm was 61.5 mm. The Threl was constant for diameters P12.5 mm. For source diameters <12.5 mm, the 41% level over-estimated the real source diameter by a factor depending on the diameter. In an in vitro set-up the SW was capable of segmenting solitary PET volumes to within 1.4 mm (1SD). For non-homogeneous signals in a clinical set-up minimal manual intervention is presently required to separate target from non-target signals. The SW may slightly underestimate target volumes when compared with CTbased volumes, but works well as a first approximation. The volume can be manually adapted to give the ultimate target volume. Conclusions: SW-based automatic delineation of the volume of 18F activity is feasible and highly reproducible. Volumes can be subsequently modified by the clinician if necessary. This approach will increase the efficiency of the planning process. c 2006 Elsevier Ireland Ltd. All rights reserved. Radiotherapy and Oncology 80 (2006) 43–50.



Keywords: FDG-PET; Functional imaging-based radiotherapy planning; PET/CT-derived target volume definition; Segmentation of PET signals; Threshold definition; Automatic 3D delineation of PET images

The metabolism of 18-fluoro-deoxy-glucose (18FDG) is proportional to the number of active cancer cells in a tumour volume. Imaging this activity by means of a positron emission tomography (PET) scan is a powerful tool in the staging of malignancies and for delineating target volumes in radiation therapy [3,5,6,29,38]. PET scanning also has the potential for removing some of the inherent uncertainties in gross tumour volume (GTV) determination. However, this aspect has not yet been exploited by radiation oncologists, nor has it been rigorously validated for this purpose. Although there is a lack of data correlating PET with spatial pathology at primary tumour sites [28], multi-modality q

Supported by the Radium Foundation, University of Zurich.



imaging is the concept of choice for target volume definition [25]. It allows the combination of optimal techniques for imaging different aspects of the anatomy with functional images, which indicate the presence of active tumour cells. Thus, functional imaging allows the planner to by-pass an essential step in defining a planning target volume (PTV). According to general procedures outlined in ICRU Report 50 [31] and Report 62 [32], ‘conventional’ target volume definition starts with the delineation of the GTV to which a margin is added, representing the possible infiltration of the tumour, to form the clinical target volume (CTV). At this stage some degree of variation has already been introduced in the procedure. Different clinicians may draw different margins as this is subject to interpretation of clinical data.

0167-8140/$ - see front matter c 2006 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.radonc.2006.07.006

44

Assessment of

18

F PET signals for target volume definition

In addition to this inter-observer variability, Giraud et al. [20] showed that the profession (radiologist, radiation oncologist or surgeon) and the level of experience of the operator also have an influence on target volume delineation. A safety margin allowing for internal organ motion and daily set-up variations is added to the CTV to form the planning target volume (PTV). This too is subject to some degree of variation. It is known that even within the context of controlled clinical trials, large variations in GTV can occur [12,19]. Deviations in the PTV can be significant. In a head and neck cancer trial Valley et al. [34] found that the ratio of the largest to the smallest PTV for the first course of radiotherapy was 6.5. For the second course the ratio was as high as 9.3. Similar differences in the target volume have been observed in lung cancer [35]. In a study of target volume definition in poorly differentiated non-small cell lung cancer, Caldwell et al. [4] have compared CT-based target volume delineations in 30 patients. They found variations in the GTV of up to a factor of 7.66. Similarly, in the dry run of a prostate trial, Kouloulias et al. [26] have found variations of up to a factor of 5 in the PTV. This lack of convergence in PTV delineations underlines the importance of quality assurance in multi-centric trials. In target volume definition, functional imaging directly defines the CTV as the infiltration pathways not visible on CT also metabolise glucose and therefore show up on the PET scan. In the study by Caldwell et al. [4], the ratio of 7.66 was reduced to 2.63 when co-registered PET data were used to delineate the GTV. In a paper on functional imaging, Bentzen [1] remarked that: ‘‘Imprecision in CTV definition remains an obstacle for high-precision RT. Functional imaging and novel biologic assays may facilitate a move from a clinical target volume to the real target volume’’. The use of PET imaging is a step in this direction. Several methods have been suggested for volume segmentation of PET signals. Automatic thresholding, developed for single photon emission computed tomography (SPECT) studies [15], has been extended to FDG PET studies in patients with lung tumours [16]. In these studies the threshold level has been

adjusted according to the signal-to-background ratio (S/B). In our study, we have used background-subtracted signals for the segmentation of PET volumes and developed a software for automatic delineation of target volumes based on a fixed relative threshold.

Materials and methods Our hypothesis is that the subtraction of background activity from PET signals only slightly influences the shape of the signal. In PET imaging, the signal is embedded in background activity. Subtracting the background results in a decrease of overall image intensity, although the signal itself is intensity and compression invariant. The shape of the signal is only affected by the image transfer function of the background at the interface between signal and background. An absolute threshold level Thabs defining the physical dimensions of the signal can be described by the formula (F) below: Thabs ¼ Bgd þ Threl ðSigmax  BgdÞ; where Sigmax is the maximum activity of the source, Bgd is the background level and Threl is the relative threshold level defined as a fixed ratio of the normalised backgroundsubtracted Sigmax (Fig. 1). Threl is the activity level that corresponds to the physical diameter of the source. This theory was verified by measurements on a combined PET/CT scanner and independently by Monte Carlo (MC) simulations. Based on the formula (F), software (SW) was generated to automatically delineate volumes of FDG PET activities. Validation of the SW was done by submitting PET images of spheres of known diameters for automatic segmentation of the volumes. Additionally, clinical PET images of a head and neck, a thymus and a breast cancer patient were used for automatic volume delineation. In the presence of nontarget signals in the vicinity of the target, the software presently needs a manual prompt to separate target from non-target signals.

Signal and threshold levels for measuring signal volume

Signal Level

10

Level corresponding to known source diameter

Sig max Th rel Th abs

Bgd 0 0

20

40

60

80

100

Position (mm) Fig. 1. Threshold levels for measuring dimensions of PET signals. For the definition of the relative threshold level Threl, the maximum signal value is set to 100% and the background value to 0%. The relative threshold level is measured from the mean background level and corresponds to the physical diameter of the source. For assessment of source diameters, the activity (MBq/ml) represented by the value of Thabs is used.

J.B. Davis et al. / Radiotherapy and Oncology 80 (2006) 43–50

Phantom measurements Small quantities of 18F water were thoroughly mixed for several minutes using a fan mixer. Two phantoms were used. The first was a locally constructed phantom consisting of an array of cylinders of inner diameter ranging from 4.5 to 26 mm and wall thickness 1 mm. These were placed between two holders inside a tank measuring 22 cm · 22 cm · 35 cm. Cylinders of diameter <4.5 mm were not used because the resolution of the PET/CT scanner is about 5 mm when used in the clinical mode. The distance between the cylinders was 30 mm. Source activities ranged from 0.01 to 0.15 MBq/ml. Background activity was simulated by filling the tank with 18F water or 18-FDG of activity within the range of 0.0005–0.007 MBq/ml. Combinations of these source and background activities provided a range of S/B ratios similar to those typically found in patients. For simplicity, they were grouped into low (S/B ratio 3:1), medium (S/B ratio 10:1) and high (S/B ratio 20:1). The second phantom, the ACR-NEMA Phantom Model ECT/NEM1 (NEMA Standards Publication NU2-1994), was used for analysis of sources of diameter >40 mm with and without background activity. This phantom consists of a Perspex cylinder of length 19.0 cm and inner diameter 19.7 cm. It contains three cylinders of inner diameter 43 mm. The cylinders and the container were individually filled with 18 F water or 18FDG providing a range of S/B ratios similar to a clinical set-up. All measurements were repeated twice each time with a slightly modified position of the phantom within the PET scanner’s field-of-view in order to have a mean of the influence of the imaging grid on image size (partial volume effect).

Image acquisition All measurements were done with a combined PET/CT scanner (Discovery LS, GE Medical Systems, Waukesha, WI). This in-line scanner is a combination of a full-ring PET scanner and a multi-slice CT scanner. Both scanners have an axial field-of-view (FOV) of 14.6 cm per single acquisition. The in-plane PET resolution is 4.8 mm full-width at half-maximum (FWHM) at the centre of the FOV. For these measurements, the same settings and parameters were used as for patients undergoing radiotherapy treatment planning. PET imaging was acquired using the 3-min per FOV mode. Images were corrected for decay of activity. Photon attenuation was corrected for by a transmission CT scan acquired using 120 kV, 120 mA and 1.0 s/rotation. All images were reconstructed using an iterative algorithm (ordered subset expectation maximization). As patients undergoing PET/CT for treatment planning purposes receive a 2D total body scan, measurements were also done in 2D mode for phantom studies. PET data were available in DICOM format.

Threshold evaluation and diameter assessment For the determination of the diameters of the cylinders, activity profiles were obtained through the centre of the signals in four directions in the transverse projection of the data. Several transverse data sets of the same cylinder were used for statistical purposes. A linear interpolation was performed between two data points of the profiles. Profile measurements spanned across the whole width of the source

45

activity plus 20 mm into the background activity region around the cylinder. The maximum background-subtracted activity level was set to 100%. The percent value of the maximum activity best representing the real diameter of each cylinder (mean of four measurements) was assessed for each cylinder. The same criteria could not be used for assessing maximum activity for all sources because large sources produced a broader and flatter profile than the smaller ones. For sources of diameter 612.5 mm, the highest pixel (4 mm · 4 mm) activity was used as maximum. For signals with a diameter P12.5 mm, the activity of a single pixel was not considered to be representative of the maximum activity of the source. Instead, the mean activity of the highest 10% adjacent pixels was set to 100%. This corresponded to two pixels for 12.5 mm, four pixels for 26 mm and nine pixels for the 43 mm cylinder. To obtain the background level, the mean value of the pixel surrounding the source at a distance of 10–20 mm from the source was set to 0%. The diameter of each cylinder was successively calculated using increasing Threl levels. The best Threl level was selected when the least square difference between the true diameter and the calculated diameter of the cylinder was reached.

Simulation of PET signals To evaluate the effect of the wall thickness on the measurement of the diameters, Monte Carlo simulations of signals and background were used. The reconstruction of a point source by an imaging system is determined by its point spread function (PSF). In an ideal image reconstruction, point sources can be added up and then convolved with a PSF. A symmetrical and constant PSF would result in a Threl of 50%. The condition for this is that signal and background are approximately constant. It was assumed that errors in quantification of the image signal could introduce an error in the imaging process which could result in a deviation from the theoretical ideal Threl of 50%. This sampling error was analysed using MC simulations. Virtual discs of activity with different S/B ratios were created. The internal resolution for the calculation was 39 lm. This is 100 times higher than the resolution of a PET image at 3.9 mm/pixel. Thus, the only error left in the calculation is the binning quantification error. First ‘‘real’’ activities of the disc with diameters ranging from 10 to 46 mm were created. The centre of each disc was shifted in the X and Y directions inside the 3.9 mm grid in random steps. The S/B ratios were 3:1, 10:1 and 20:1. The resulting image was convolved with a Gaussian filter with a FWHM of 8 mm resulting in a blurred image border. The resolution of this image was then reduced to 3.9 mm pixel size as available in the PET scanner. To minimize quantification errors, an up-sampling by a factor of 2 with bi-linear filtering is applied to the ‘‘PET image’’ yielding an image with 1.95 mm pixel size. A region-growing algorithm then searches for the best threshold. This results in a contour which optimally coincides with the original disc.

Description of software As for the simulation of the PET images, an up-sampling of a factor of 2 with tri-linear filtering is applied to the clinical PET signals in order to minimize errors in sampling and

46

Assessment of

18

F PET signals for target volume definition

quantification. A first point (P1) is manually selected within the target region. This serves as a starting point to estimate the maximum signal (Sigmax_est). A second point (P2) is selected in the background region (Bgd). From these two points, an estimated absolute threshold Thabs-est is calculated based on formula (F). A region-growing algorithm starts from P1 and produces a preliminary target volume based on Thabs-est, but this may also include non-target structures (Fig. 2a). To discard these, a third point (P3) is selected within the non-target structure. Two region-growing algorithms are started, one from Sigmax_est and one from P3. The two regions are allowed to grow simultaneously and the algorithm is stopped before they begin to merge into one. Separation of the two regions is done using a Voronoi algorithm [24,37]. The algorithm uses a watershed-like principle to achieve the geometric separation of target and nontarget volumes. Fig. 2b shows the separated regions. Only those voxels which are inside the target and lie within half of the FWHM of the PET image resolution from Thabs are included in the final assessment of the true mean maximum signal level Sigmax_mean. A similar procedure is adopted for the Bgdmean assessment. The final value of Thabs used for the determination of target volume is then calculated using formula (F), yielding Threl. Finally, a region-grower based on Threl produces the region ‘‘Target’’. Fig. 2c shows the final target volume. The procedure can be repeated for subsequent target volumes.

Validation of software The locally constructed phantom used for cylinder measurements was also used for in vitro validation of the software. It was re-fitted with spheres of known diameters and filled with 18F activities producing similar S/B ratios as for the cylinder measurements. The phantom was imaged using the same procedure. The software was then used to generate the volumes of the activities based on the calculated Threl level.

In vivo validation was done in a retrospective analysis of PET/CT data. The SW was used on PET data from PET/CT data previously used for treatment planning purposes. GTVs automatically generated by the SW were compared with GTVs delineated by a clinician.

Results Threshold level assessment For the locally constructed phantom, the smallest source diameter evaluated was 4.5 mm. For these diameters, no histogram of the source activity could be generated and the signal presented visually as a partial volume irrespective of its position within the image grid. Despite this, data for sources of diameter <12 mm were not discarded from the analysis. The largest cylinder diameter in this phantom was 26 mm. Data obtained with the ACR NEMA phantom were used to assess Threl for sources of diameter 43 mm and the influence of the background activity on the measurements. Monte Carlo simulations of the effect of the cylinder walls showed that the mean relative threshold Threl of 44 ± 2% corresponded to known disc diameters >16 mm. Threl was influenced by 6% and 1.7% for a S/B of 3:1 and 10:1, respectively. The influence of the wall thickness for S/B of 20:1 was negligible. Correcting for this in the measurements, the source diameter which best corresponded to the physical diameter of the cylinder was obtained at a mean Threl of 41 ± 2.5% at the 95% confidence level (Fig. 3). Only cylinders with a diameter of 12 mm and larger were used for the evaluation.

Validation of software Results of the in vitro validation are summarized in Fig. 4. For solitary spheres of diameter >15 mm the accuracy of the SW was 6±1 mm in the transverse plane and depended to some extent on the relative position of the

Fig. 2. Description of the software using a clinical example. (a) One starting point for the target and one for background are manually selected. A first estimate of the target volume is made. This may include a non-target region. In this example, it is part of the brain. (b) The target volume is retrieved by starting a separation Voronoi algorithm which divides the data into a target region and a non-target region. (c) The final calculation of the tumour volume is based on voxel information within the target Voronoi region only. Threl is solely based on the mean value of the signal and the background region immediately surrounding the signal.

J.B. Davis et al. / Radiotherapy and Oncology 80 (2006) 43–50

Relation between Th rel and known cylinder diameters for different S/B

1.00

Relative Threshold Th rel

47

S/B 3:1 S/B 10:1 S/B 20:1 Th 0.41

0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0

4

8

12

16

20

24

28

32

36

40

44

Cylinder Diameter (mm)

Measured Diameter (mm)

Fig. 3. Relative threshold (Threl) level assessed for a range of known cylinder diameters and S/B ratios. The least square fit method was used. The level is constant for source diameters down to 12mm. The level best corresponding to the known diameter of the cylinder was obtained at a mean Threl of 41% ± 2.5. Only a minimal dependence on S/B ratio was observed, except for the lowest S/B ratio of 3:1, but this might be due to the low signal level.

Measured to known diameter of spheres (based on 41 Th rel)

40 35 30 25 20

S/B 3:1 S/B 10:1

15

S/B 20:1

10 10

15

20

25

30

35

40

Sphere Diameter (mm) Fig. 4. In vitro software validation using spheres of known diameter. There is good agreement between the measured and the known diameters within the range of 16–34 mm at the 41% Threl level. Measurements for objects having a S/B ratio of 3:1 have a larger SD than for those with a S/B ratio of 10:1 or a 20:1.

sphere within the FOV. The SD of the data for the S/B ratio of 3:1 is 2.8% and is larger than for the data of the 10:1 S/B ratio (1.2%) and of the 20:1 S/B ratio (1.7%). These results indicate that although the basic measurements were made in a 2D setting, the software works equally well in 3D. The accuracy in the longitudinal plane is limited by the slice thickness and distance from the central axis of the FOV. For the set-up used, the accuracy was <3 mm. Clinical validation of the SW was done by a retrospective comparison of SW-derived PET-based target volumes with clinician-derived CT-based target volumes. Three examples are shown in Fig. 5. These include the head and neck patient used in Fig. 2 to demonstrate the automatic outlining procedure of the SW. The two target volumes are similar. The PET-based volume can be readily modified by the clinician if necessary. In Fig. 5b a breast tumour is shown. The two volumes are a perfect match and no manual adjustment is necessary. Finally, a thymoid tumour is shown Fig. 5c. Here the CT-based volume and the PET-based volumes are similar in size and in position, but in places the PET-based volume reaches out of the CT-based volume. Whether this is due to tumour cells extending beyond the clinician-derived volume or due to the motion of heart and lung is unknown.

Discussion The major impact of PET imaging in oncology lies in staging and re-staging for a broad variety of tumours [9]. Present integrated PET/CT systems achieve a sufficiently high sensitivity and specificity to allow appropriate staging of nodal disease, for example in lung cancer [17,21,27,33]. Conceptual changes in radiotherapy planning incorporating PET in staging of cancer disease have been observed in about 15–20% of patients undergoing evaluation for curative radiotherapy [8,14,17,36]. In two dose escalation studies of non-small cell lung cancer (NSCLC), De Ruysscher et al. [13] and Holloway et al. [23] have shown that a reduction of the target volume may be attempted if PET information is incorporated into the planning process. However, Lavrenkov et al. [28] rightly state that PET should not be used as the only source of information in NSCLC and Gregoire [22] generally calls for caution when using PET images for treatment planning. Large mismatches between CTand PET-derived volumes have been observed by Geets et al. [18], but they note that [11C]methionine-PET volumes in pharyngo-laryngeal squamous cell carcinoma are similar to CT-derived target volumes. They point out, however, that one of the strengths of the addition of PET data to

48

Assessment of

18

F PET signals for target volume definition

Fig. 5. Three clinical examples of target volume definitions. The SW-derived PET-based target volume is shown in red and the physician-drawn CT-based volume in green. (a) A head and neck carcinoma. The PET-based volume is slightly smaller than the CT-based volume but they are never-the-less quite similar. (b) An example of a relapsing breast cancer. Here, an almost perfect match can be observed. (c) A case of a thymus tumour. The PET-based volume is similar to the CT-based one, except that at places the PET-based volume reaches out of the CT-based volume. This could be due to tumour cells extending beyond what the clinician can see on a CT scan or motion of the heart and lung.

the planning process lies in the difference between the PET and CT data. In this study, we have used a background-subtracted fixed threshold method to quantify the dimensions of PET signals. The threshold level of 41% is in agreement with our previous clinically assessed threshold of 40% for anal cancer as indicated by Ciernik et al. [7]. Ours is effectively a flexible threshold method because subtracting background activity, which may not necessarily be constant, will affect the absolute threshold level which defines the volume of activity. Thus, we have attempted to find a universal approach to quantify 18F PET signals for the purpose of standardizing target volume definition. We have then designed a software to automatically derive a target volume. Standardizing target volumes using PET signals could be crucial also for prospective collaborative studies, if trials with specific radiotherapy questions are based on the biological activity of tumours. Variable biological characteristics, such as tumour growth, perfusion or re-perfusion, replication and hypoxia, may then be specifically investigated with adapted targeting opening the way for optimised radiation dose and fractionation schedules. Erdi et al. [15] investigated the volumetric accuracy of SPECT signals in vitro. They state that a fixed threshold may result in limited volume estimate accuracy. Instead, they suggest a threshold based on tumour size and image contrast. The issue of correct volumetric analysis of PET signals is becoming more pertinent in radiotherapy, as PET data are becoming more readily available for treatment planning. Black et al. [2] have compared a linear regressive threshold method with a constant threshold method (42% of maximum intensity) and suggest that both models could be improved. Daisne et al. [11] proposed a segmentation algorithm for three-dimensional PET structures based on the S/B ratio providing useful volume estimates without knowledge of the investigated structure. The method defines the threshold of the signal of interest with respect to the background activity present. Their model, where y is the

threshold of activity Thresholdactivity (percentage of maximum activity), is represented by: Thresholdactivity ¼ a þ b=ðS=BÞ; can be easily converted to our model of Thabs ¼ Thresholdactivity ¼ Sigmax  ½Threl þ ð100  Threl Þ=ðS=BÞ; if a ¼ Threl and b ¼ 100  Threl : Calculating Threl for their data, we obtain threshold levels of 26.1 ± 2.5%, 23.0 ± 2.1%, 29 ± 3.1% and 32.4 ± 3.2%. This clearly indicates a dependence on the image reconstruction algorithm and hence only applies to a specific algorithm and PET scanner. It might therefore be advantageous to have a standard phantom and software for the calculation of Threl levels. Nestle et al. [30] have investigated four different methods of PET-based GTV delineation in 25 cases of NSCLC and compared them with CT-based data. These methods were: (1) GTVvis – by visual determination of the GTV, (2) GTV40 – using a threshold at 40% of the Standardized Uptake Value SUVmax, (3) GTV2.5 – using an iso-contour of SUV2.5 around the tumour, and (4) GTVbgd – using a threshold of (0.15 · Imean, the mean maximum intensity) + bgd. They found substantial differences in the resulting volumes. As in our study, they found that the background-subtraction method, the GTVbgd, led to the most satisfactory volumes. The data from Yaremko et al. [39] also fit our model. Their data from the detection of PET-derived spherical volumes are in good agreement with our data. Normalising the signal to 1 the data can be described by: Thresholdactivity ¼ Bgd þ Threl ð1  BgdÞ ¼ Threl þ Bgdð1  Threl Þ ¼ 0:35 þ 0:65Bgd; which is another way of writing formula (F), and yields a similar result. The relative threshold level here is 35%,

J.B. Davis et al. / Radiotherapy and Oncology 80 (2006) 43–50

which is similar to ours bearing in mind that these results are obtained from the steep part of the curves. Smaller spheres are affected by the partial volume effect. The partial volume effect observed in cylinders of diameter <12 mm could be corrected for in our model. After determining the Threl level, this value can be corrected by the partial volume effect and then increase the Thabs level until the correct volume is reached. However, it is doubtful if target volumes of diameter <12 mm can be accurately determined with a system whose spatial resolution is 4.8 mm FWHM without further improvement of the reconstruction algorithm and further manipulations of the software. A volume represented by a single voxel can give the same result as a volume comprising four adjacent voxels but of lower activity. Data available indicate that for diameters <12 mm a flexible threshold may be necessary to predict the real diameter. This tends to support the conclusions reached by Erdi et al. [15]. The matching accuracy between PET and CT is especially important for small volumes. Daisne et al. [10] have evaluated the matching accuracy of their co-registered PET/CT and found discrepancies of several mm. The techniques employed in the development of the algorithm presented here are simple but have been shown to work well even where signal and background are not homogeneous. However, in this study the segmentation of clinical PET images is based on a threshold determined on stationary volumes in vitro. The question of tumour movement has not been addressed. Because of the long acquisition time for a PET study, clinical images inevitably contain movement artefacts. The SW is presently being assessed clinically and preliminary results show that a realistic target volume can be delineated. If necessary, this can be optimised by the clinician to produce the ultimate target volume. In any case this is faster than delineating a series of 2D outlines on transverse CT data sets and is an operator-independent target volume. The procedure can be repeated for subsequent target volumes.

Conclusions In this study, a new approach to target volume definition which will help reduce operator-dependent variations has been described. The model presented is intuitive, robust and easy to use. It can be directly implemented in the planning process without additional tools. For target volume definition, the algorithm can be used clinically but for the final volume delineation CT or MRI data must be taken into account. Its major advantage lies in the potential that PETbased treatment planning can provide a standardized first line target volume. This approach leaves the clinician with a working volume, which may be ultimately adapted in a second step. Such a procedure may speed the planning process significantly, reduce time delays during planning, allow a more rapid start of treatment and reduce patient waiting time and distress. For tumours located in high background or in heterogeneous activity regions, or for those in the proximity of high activity regions, the SW presently needs an additional manual intervention to differentiate between target and non-

49

target structures. As for tumour heterogeneity, at least for the near future, GTV delineation will still greatly depend on an input from the clinician. * Corresponding author. Beatrice Reiner, Radiation Oncology, University Hospital, Raemistrasse 100, 8091 Zurich, Switzerland. E-mail address: [email protected] Received 16 June 2005; received in revised form 18 April 2006; accepted 6 July 2006

References [1] Bentzen SM. High-tech in radiation oncology: should there be a ceiling? Int J Radiat Oncol Biol Phys 2004;58:320–30. [2] Black QC, Grills IS, Kestin LL, et al. Defining a radiotherapy target with positron emission tomography. Int J Radiat Oncol Biol Phys 2004;60:1272–82. [3] Bradley J, Thorstad WL, Mutic S, et al. Impact of FDG-PET on radiation therapy volume delineation in non-small-cell lung cancer. Int J Radiat Oncol Biol Phys 2004;59:78–86. [4] Caldwell CB, Mah K, Ung YC, et al. Observer variation in contouring gross tumor volume in patients with poorly defined non-small-cell lung tumors on CT: the impact of 18FDG-hybrid PET fusion. Int J Radiat Oncol Biol Phys 2001;51:923–31. [5] Calvo FA, Domper M, Matute R, et al. 18F-FDG positron emission tomography staging and restaging in rectal cancer treated with preoperative chemoradiation. Int J Radiat Oncol Biol Phys 2004;58:528–35. [6] Ciernik IF, Dizendorf E, Baumert BG, et al. Radiation treatment planning with an integrated positron emission and computer tomography (PET/CT): a feasibility study. Int J Radiat Oncol Biol Phys 2003;57:853–63. [7] Ciernik IF, Huser M, Burger C, Davis JB, Szekely G. Automated functional image-guided radiation treatment planning for rectal cancer. Int J Radiat Oncol Biol Phys 2005;62:893–900. [8] Cohade C, Osman M, Leal J, Wahl RL. Direct comparison of 18FFDG PET and PET/CT in patients with colorectal carcinoma. J Nucl Med 2003;44:1797–803. [9] Czernin J, Phelps ME. Positron emission tomography scanning: current and future applications. Annu Rev Med 2002;53:89–112. [10] Daisne J-F, Sibomana M, Bol A, Cosnard G, Lonneux M, Gregoire V. Evaluation of a multimodality image (CT, MRI and PET) coregistration procedure on phantom and head and neck cancer patients: accuracy, reproducibility and consistency. Radiother Oncol 2003;69:237–45. [11] Daisne J-F, 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;69:247–50. [12] Davis JB, Reiner B, Dusserre A, Giraud J-Y, Bolla M. Quality assurance of the EORTC Trial 22911. A phase III study of postoperative external radiotherapy in pathological stage T3N0 prostatic carcinoma: The dummy run. Radiother Oncol 2002;64:65–73. [13] De Ruysscher D, Wanders S, Minken A, et al. Effects of radiotherapy planning with a dedicated combined PET-CTsimulator of patients with non-small cell lung cancer on dose limiting normal tissues and radiation dose-escalation: a planning study. Radiother Oncol 2005;77:5–10. [14] Dizendorf EV, Baumert BG, von Schulthess GK, Lutolf UM, Steinert HC. Impact of whole-body 18F-FDGPET on staging and managing patients for radiation therapy. J Nucl Med 2003;44:24–9. [15] Erdi Y, Wessels B, Loew M, Erdi A. Threshold estimation in single photon emission computed tomography and planar

50

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

Assessment of

18

F PET signals for target volume definition

imaging for clinical radio-immunotherapy. Cancer Res 1995;55:5823s–6s. Erdi YE, Mawlawi O, Larson SM, et al. Segmentation of lung lesion volume by adaptive positron emission tomography image thresholding. Cancer 1997;80:2505–9. Erdi YE, Rosenzweig K, Erdi AK, et al. Radiotherapy treatment planning for patients with non-small cell lung cancer using positron emission tomography (PET). Radiother Oncol 2002;62:51–60. Geets X, Daisne J-F, Gregoire V, Hamoir M, Lonneux M. Role of 11-C-methionine positron emission tomography for the delineation of the tumor volume in pharyngolaryngeal squamous cell carcinoma: comparison with FDG-PET and CT. Radiother Oncol 2004;71:267–73. Giraud JY, Dusserre A, Conil M, Davis JB, Bolla M. Results of the dummy-run of EORTC trial no.22961: a phase III study of long term hormonal treatment on locally advanced protatic carcinoma treated by radiation therapy and a 6 months CAB. Int J Radiat Oncol Biol Phys 2001;51:378. Giraud P, Elles S, Helfre S, et al. Conformal radiotherapy for lung cancer: different delineation of the gross tumor volume (GTV) by radiologists and radiation oncologists. Radiother Oncol 2002;62:27–36. Gould MK, Kuschner WG, Rydzak CE, et al. Test performance of positron emission tomography and computed tomography for mediastinal staging in patients with non-small-cell lung cancer: a meta-analysis. Ann Intern Med 2003;139:879–92. Gregoire V. Is there any future in radiotherapy planning without the use of PET: unravelling the myth. Radiother Oncol 2004;73:261–3. Holloway CL, Robinson D, Murray B, et al. Results of a phase I study to dose escalate using intensity modulated radiotherapy guided by combined PET/CT imaging with induction chemotherapy for patients with non-small cell lung cancer. Radiother Oncol 2004;73:285–7. Kenneth E, Hoff I, Culver T, Keyser J, Lin M, Manocha D. Fast computation of generalized Voronoi diagrams using graphics hardware. In: Proceedings of the sixteenth annual symposium on computational geometry; 2000. p. 375–6. Kessler ML, Pitluck S, Petti P, Castro JR. Integration of multimodality imaging data for radiotherapy treatment planning. Int J Radiat Oncol Biol Phys 1991;21:1653–67. Kouloulias VE, Giraud J-Y, Davis BJ, Dusserre A, Zurlo A, Bolla M. Quality assurance of the 22961 EORTC trial. A phase III study of the optimal combination of hormonal adjuvant treatment by LHRH analogue and radiation therapy for the management

[27]

[28]

[29]

[30]

[31]

[32]

[33] [34]

[35] [36]

[37]

[38]

[39]

of locally advanced prostate cancer: the dummy run. Radiother Oncol 2004;73:11–20. Lardinois D, Weder W, Hany TF, et al. Staging of non-smallcell lung cancer with integrated positron-emission tomography and computed tomography. N Engl J Med 2003;348:2500–7. Lavrenkov K, Partridge M, Cook G, Brada M. Positron emission tomography for target volume definition in the treatment of non-small cell lung cancer. Radiother Oncol 2005;77:1–4. Lee YK, Cook G, Flower MA, et al. Addition of 18F-FDG-PET scans to radiotherapy planning of thoracic lymphoma. Radiother Oncol 2004;73:277–83. 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;46:1342–8. Prescribing, recording and reporting photon beam therapy, ICRU Report, 50. Bethesda, MD: The International Commission on Radiation Units And Measurements; 1993. Prescribing, recording and reporting photon beam therapy, ICRU Report, 62. Bethesda, MD: The International Commission on Radiation Units And Measurements; 1999. Townsend DW, Carney JP, Yap JT, Hall NC. PET/CT today and tomorrow. J Nucl Med 2004;45:4S–14S. Valley J-F, Bernier J, Tercier P-A, et al. Quality assurance of the EORTC radiotherapy trial 22931 for head and neck carcinomas: the dummy run. Radiother Oncol 1998;47:37–44. Valley JF, RO M. Comparison of treatment techniques for lung cancer. Radiother Oncol 1993;28:168–73. Vanuytsel LJ, Vansteenkiste JF, Stroobants SG, et al. The impact of 18F-fluoro-2-deoxy- glucose positron emission tomography (FDG-PET) lymph node staging on the radiation treatment volumes in patients with non-small cell lung cancer. Radiother Oncol 2000;55:317–24. Voronoi GM. Nouvelles Applications des parame `tres continus a ` la the ´orie des formes quadratiques. Deuxie `me Me ´moire: Recherches sur les paralle ´loe `dres primitifs. J Reine Angew Math 1908;134:198–287. Vrieze O, Haustermans K, Wever WD, et al. Is there a role for FGD-PET in radiotherapy planning in esophageal carcinoma? Radiother Oncol 2004;73:269–75. Yaremko B, Riauka T, Robinson D, Murray B, McEwan A, Roa W. Threshold modification for tumour imaging in non-small-cell lung cancer using positron emission tomography. Nucl Med Commun 2005;26:433–40.