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

MATERIALS AND METHODS

Data of 25 patients with histologically proven primary NSCLC were used, who had a routine ¹⁸F-FDG PET examination for staging purposes. All patients had an option for radiotherapy at the time of the PET examination, although not all patients finally received this treatment.

Tumor stages were T1 or T2 in 17 patients and T3 or T4 in 8 patients. Eighteen patients had positive mediastinal nodes; 6 patients had distant metastases.

Our investigation focuses on the GTVs concerning the primary tumors, regardless of the N and M stages.

After obtaining informed consent, patients underwent routine whole-body ¹⁸F-FDG PET (250 MBq ¹⁸F-FDG; fasting blood glucose level, <150 mg/dL; CTI/Siemens ECAT ART PET scanner; 6 or 7 bed positions; attenuation correction by transmission scanning with ¹³⁷Cs single-photon transmission; axial spacing 3.4 mm; iterative reconstruction into 128 × 128 pixels of 5.1 mm), with the acquisition being started 90 min after injection. Afterward (160 min after injection), an additional PET scan of the chest was acquired in radiotherapy treatment position (2 or 3 bed positions). On the same day, spiral planning CT of the chest (El-Scint TWIN FLASH CT; 3-mm slice thickness, 512 × 512 pixels of 0.98 mm, flat breathing) was performed in identical position verified by laser localizer, skin marks, and photographic documentation. The coregistration of CT and PET data (23) was performed by a Hermes (Nuclear Diagnostics) workstation; the data were then transferred to the radiotherapy planning system (Philips Pinnacle).

As a first step, in all patients, an experienced double board-certified nuclear medicine and radiotherapy physician used the region-of-interest (ROI) standard evaluation tool provided by the manufacturer of the PET system and a global logarithmic scaling to generate a “visual” PET GTV, comprising the tissue considered visually as part of the malignant primary tumor (GTV_vis). Clinical information and CT reports of the patients were used in this process but CT images or image fusion was not used.

Then, for all tumors, 2 further GTVs were defined at the PET console. ROIs were positioned around the tumors slice by slice in the volume file, using first an isocontour of SUV = 2.5 (GTV_2.5) and, second, an isocontour of 40% of the SUV_max of the whole lesion (GTV₄₀) similarly for all slices.

In our opinion, the radiotherapy planning system (PS) is the most likely place for PET target volumes to be defined in clinical practice. The data transfer described converts the voxel values of PET activity (kBq/mL) to visual intensities (I), so that the SUV is lost. Furthermore, the matrix is changed from 128 × 128 to 512 × 512. Therefore, the delineation of the fourth set of GTVs was done on the PS using an in-house algorithm.

As reported earlier in part (24), in-house phantom measurements were performed using spheres with varying diameters and source-to-background activities.

The physical and mathematic features and results of these phantom experiments will be published separately in more detail. However, the clinical application of the resulting algorithm was included into our comparison.

In short, by means of the PS, thresholds for volume contouring are determined by a function of tumor and background intensities: I_threshold = (0.15 × I_mean) + I_background.

In patient datasets, I_mean was calculated as the mean intensity of all pixels surrounded by the 70% I_max isocontour within the tumor. The rationale for the choice of I_mean rather than I_max was to minimize the influence of statistically not representative maximum values on the resulting threshold.

I_background was defined as follows: Anatomic entities adjacent to the tumor (e.g., lung, mediastinum, liver) were identified. By visual comparison of these, the structure with the highest ¹⁸F-FDG uptake was defined as “relevant backround.” A ROI was placed into the relevant background structure at a safe distance from the target, and the mean SUV of this ROI was used as I_background for threshold calculation. This procedure was established on the assumption that, for tumors adjacent to various anatomic structures, those with more intense ¹⁸F-FDG accumulation after injection (e.g., mediastinum, liver) were more relevant for threshold calculation than faintly accumulating tissue (e.g., lung). We further assumed, and confirmed this assumption by exploratory measurements, that normal organs show a rather homogeneous ¹⁸F-FDG accumulation after injection within themselves.

Applying the resulting thresholds in the 22 patients eligible for evaluation at the PS (in 3 patients, data transfer failed because of technical reasons), tumor contours were outlined automatically and then manually corrected to exclude nontumor tissue—for example, myocardium. This procedure led to the fourth set of GTVs (GTV_bg).

As a common feature in lung cancer, in some cases ¹⁸F-FDG–positive lymph nodes directly adjacent to the primary tumors could not be separated from the tumor itself. Here, the whole structure accumulating ¹⁸F-FDG was included into all GTVs as if it was part of the primary tumor.

To correct for a possible influence of the tumor size on the differences detected, virtual spheric radii were calculated for all GTVs (R_vis, R₄₀, R_2.5, R_bg, R_CT). In addition, this yielded a value comparable to clinical practice, where the differences in radius would reflect the distance between the contours drawn in the same image.

In CT, as common in lung cancer, in most tumors the circumferences could only be partially contoured unequivocally. However, in 5 patients, a peripheral tumor was fully delimitable. Here a GTV derived from CT was generated using the soft-tissue window with respect to lung window (GTV_CT). To correct for breathing excursions during the PET scan, and therefore to provide a measure for the size of PET GTVs to be expected, “expanded CT volumes” were calculated (GTV_CTexp). Following the lower levels of tumor movements reported, and of correction margins recommended in the literature for radiotherapy planning (25–28), the expansion was 0.15 cm lateral, 0.2 cm anteroposterior, and 0.3 cm craniocaudal. Radius values derived from the unexpanded CT volumes were “expanded” by 0.25 cm.

By means of visual characterization of the ¹⁸F-FDG accumulation, the tumors were classified as “rather homogeneous” or “grossly inhomogeneous.”

The results were evaluated by standard methods of descriptive statistics, including combined t test and the Pearson correlation.

RESULTS

The SUV_max values of the primary tumors were in mean 17.1 (range, 1.7–38.7). As expected (29), this value was significantly higher compared with the whole-body examinations 90 min after injection (mean SUV_max, 13.4; P < 0.0001). However, 3 patients showed a decrease of SUV_max. One of these had the overall minimum SUV_max of 1.7 after an initial value of 3.6. Because the patient had a malignant lesion, later confirmed histologically, we decided to include this case as far as possible into the present investigation. However, no GTV_2.5 could be determined.

Table 1 shows the results with respect to volumes and spheric radii of the GTVs created as well as the results of the statistical comparison. Mean GTV_vis was 157.7 mL, representing a R_vis of 3.03 cm. As can be seen, there are clear differences between the GTVs created. While the differences between GTV_vis and GTV_2.5 appear rather small, all other differences were equal to or larger than the spatial resolution of the PET system. Despite the small group of patients examined, this was statistically highly significant.

The maximum difference in radius detected in an individual patient was 2.22 cm (R_vis − R₄₀), 41% of R_vis of this tumor.

The differences between GTV_vis and GTV_2.5 compared with GTV₄₀ or GTV_bg correlated significantly with the SUV_max, the size of the lesion, and the presence of gross inhomogeneity (all P values < 0.01).

On further examination of these findings, 2 subgroups were analyzed exploratively:

In 8 patients, the tumors showed a grossly inhomogeneous ¹⁸F-FDG accumulation. These tumors (Table 2) were significantly larger than the others (mean R_vis = 4.16 vs. 3.03 cm; P < 0.0001). However, they did not show a significantly different SUV_max (mean, 17.1 vs. 19.7; not significant [NS]).

In this group, the differences were as before but were more pronounced. Furthermore, by visual impression, in at least 3 of the 8 patients, there was grossly inadequate coverage of the malignant tissue by the GTV₄₀ (Figs. 1 and 2), whereas GTV_bg proposed a better concordance of the ¹⁸F-FDG accumulation with the lesions depicted by CT.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1.

Two slices of image fusion in patient 1 with inadequate tumor coverage by GTV₄₀ (red) in ¹⁸F-FDG–inhomogeneous tumor (green = outline of GTV_bg).

The 5 tumors, which were fully delimitable by CT (Table 3), were all located peripherally and, on average, smaller than those of the whole group examined (mean GTV_vis = 66.5 mL). The differences between the GTVs here were less pronounced than seen before (Fig. 3).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 2.

Planning CT scan (A) and corresponding fusion image (B) of patient 8 show inadequacy of GTV₄₀: green = GTV_bg, red = GTV_CT, yellow = GTV₄₀.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 3.

Example of difference between target volumes in patient 4 with a tumor clearly delimitable by CT. (A) ¹⁸F-FDG PET. SUV_max = 30. Isocontours: narrow = GTV₄₀, wide = GTV_2.5. (B) Corresponding planning CT. Isocontours: red = GTV₄₀, green = GTV_bg, yellow = GTV_CT.

Despite the small number of cases, the GTV_vis, GTV_2.5, and GTV_bg values correlated clearly with GTV_CT (correlation coefficient = 0.96–0.98; all P values ≤ 0.02), whereas the GTV₄₀ did not (correlation coefficient, 0.70; P = NS). However, because of the small number of cases, this result must be regarded with caution, and further statistical evaluation was not done.

DISCUSSION

The present investigation attempts to contribute to the discussion about the standardization of target volumes in radiotherapy planning as derived from ¹⁸F-FDG PET. Though there are numerous investigations on phantom measurements addressing this problem, only a small number of studies including patient data have been published (16,17,30). To our knowledge, there is no clinical investigation on the potential differences between the various approaches.

Our study addressed this question by investigating the primary tumors of patients with NSCLC. Because of substantial anatomic and pathohistologic differences between primary tumors and lymph nodes, we believed that these should be dealt with separately.

Primarily, the aim of our investigation was to detect and quantify any differences between the delineation philosophies for ¹⁸F-FDG PET. During this comparison, we perceived the need to determine the “true” volumes of the lesions investigated. However, for patient data—in contrast to phantom measurements—there is no golden standard for the evaluation of volumes as measured by different imaging modalities except pathologic specimens, which were not available because of the nature of our patient population. Furthermore, in lung tumors, results of in situ volumetry will always be dependent on how the individual method deals with tumor motion. Despite blurring, volumes of chest tumors as measured by ¹⁸F-FDG PET (representing the accumulation averaged over several breathing cycles) would be expected to be equal or larger than the volumes as measured by CT, representing rather a “snapshot” of the density at one point of time during the breathing cycle (28). We therefore calculated “expanded” CT volumes according to the smallest margins recommended (25–28) for motion correction in radiotherapy planning for tumors clearly delimitable by CT. To us, these expanded volumes appeared to be closest to the true PET volumes to be expected. For the other tumors, which—as frequently observed in lung cancer—were not fully delimitable, and, therefore, for which no CT volume could be determined, we compared the PET GTVs visually to CT findings in fusion images. As illustrated in Figures 1 and 2, even in tumors that are not fully delimitable, it can at least be determined whether or not a mass is roughly surrounded by an isocontour.

In the overall comparison of the 4 philosophies of contour definition, we found significant differences between the resulting target volumes. These differences correlated with SUV_max, lesion size, and tumor inhomogeneity.

Visual definition and the application of a constant isocontour of SUV = 2.5 rendered surprisingly similar results. No differences exceeding the spatial resolution of the PET scan were observed between these 2 methods. However, visual judgment is very much dependent on the individual investigator and display window setting—for example, the type of gray scale applied. Furthermore, the GTVs generated by these approaches appeared rather large in comparison with expanded CT data. Exploratory results on mediastinal lymph nodes have further shown us that, for faintly accumulating structures in a background with relatively high activity, the SUV = 2.5 isocontours are not suitable for target volume delineation.

The most striking findings of this study relate to the 40% SUV_max approach, which, in general, yielded the smallest set of GTVs. The differences in comparison with the visual or the SUV_2.5 approach were found to be larger than the resolution of the PET system.

In patients with inhomogeneous tumors, we observed a 3.6-fold difference in mean volume (GTV₂₅ − GTV₄₀), corresponding to differences in radius up to 2.2 cm. In 3 of 8 patients, the visual impression of inadequate coverage of the malignant tissue by the 40% isocontour was obvious (Figs. 1 and 2).

In the tumors fully delimitable by CT, there was no correlation of GTV₄₀ with GTV_CT, though this was the case for all other concepts.

The mean GTV₄₀ in the present investigation was in the range of GTV_CT without expansion in well-delineated tumors, which were relatively small and situated peripherally. This finding is in line with the data by Erdi et al. (16), who used the 40% approach for GTV delineation. This group has meanwhile developed systems of breath control to be used for planning and irradiation (9,31), thus avoiding the problem of incongruent imaging of tumor motion. However, if breath control is not applied, “tailoring down” CT-defined GTVs to PET contours generated by an automatic 40% SUV approach in lung cancer does not safely lead to a complete coverage of the malignant tissue.

The philosophy of defining target volume contours by an algorithm with respect to local background was proposed by several authors (17,21). One group even showed superiority of the ¹⁸F-FDG PET volumes defined by a source-to-background algorithm over CT- and MRI-measured volumes when compared with histologic specimens in larynx tumors (30).

The algorithm developed in our institution, to be used in the PS to define the threshold for the PET-positive volume as a function of the intensities of tumor and background, led to volumes (GTV_bg) of an intermediate size between the GTV_2.5 and the GTV₄₀. In visual comparison with CT data, these volumes seemed to fit pathoanatomic structures better than the GTV₄₀.

The GTV_bg algorithm is closely related to the departmental setup. The exploratory use of the formula for volume definition away from the PS—for example, at the PET system itself—led to significantly different volumes. It must be pointed out that technical and software factors of all steps involved do have an important influence on the structure of image data and on resulting volumes. Therefore, any contouring algorithm must be regarded as system specific for use at the point of the radiotherapy chain for which it was developed. As with all other parts of the radiotherapy chain, PET contouring algorithms must be quality controlled for each system, including phantom measurements, before being used in any application.

The GTV_bg method, however, appears to be more stable against the inhomogeneity of tumor uptake, and the broad variation of SUV_max values between patients, than, for example, the 40% approach.

An important issue to be discussed before the use of complex algorithms in radiotherapy planning for lung cancer is the choice of the “relevant” background. Until now, most publications in this context have been phantom studies (21,22) so that the current literature does not provide solutions for this problem. To our knowledge, the only investigation published to date on the use of a source-to-background algorithm in patients (30) focused on larynx tumors. In the head-and-neck region, the differences in normal tissue accumulation are not as high as in the thorax. In the chest, mean ¹⁸F-FDG uptake in normal tissues may vary between a SUV of <1 (lung) up to a SUV of >3 (liver). Depending on the algorithm used, these differences may lead to significantly different thresholds, especially in contouring tumors with only faint accumulation of ¹⁸F-FDG. For the present investigation we used a differentiated approach, choosing and measuring the relevant background as defined for each patient separately, with the encouraging results reported.

A possible limitation of the present investigation might be found in the late acquisition of the planning PET scans. Other groups have reported on PET scans acquired much earlier after injection (e.g., 45–60 min). It is known that the ¹⁸F-FDG uptake in malignant tumors rises over time, though decreasing in other tissues (29,32,33). This may possibly lead to an accentuation of our findings. However, exploratory delineation of GTVs both in early and in late PET scans of several patients did not show relevant changes of the results.

Overall, there is a great need for imaging methods that precisely depict tumor tissue to aid the delineation of target volumes in high-dose 3-dimensional irradiation. Because of the high image contrast, and the comparably high diagnostic accuracy, ¹⁸F-FDG PET has a large potential in this context, which urgently needs to be integrated into clinical trials.

A first prospective study has already shown that the probability of local tumor recurrence outside the planning target volume is low after irradiating only the ¹⁸F-FDG–positive tissue (34). It is clear that patients with large—and therefore inhomogeneous—tumors might benefit from dose escalation (35). Because the differences between the philosophies for target volume definition by ¹⁸F-FDG PET are most pronounced in this group of patients, the development of a standard for the delineation of ¹⁸F-FDG–positive tissue is needed.

CONCLUSION

The different techniques used for tumor contour definition by ¹⁸F-FDG PET in radiotherapy planning resulted in substantially different volumes, especially in patients with inhomogeneous tumors.

Because of possibly incomplete tumor coverage, to us, the 40% SUV_max concept does not appear generally suitable for target volume delineation unless systems are used for breath control.

More complex algorithms—for example, contrast-oriented methods for contour definition—should further be evaluated with special respect to patient data.

It must be emphasized that such algorithms are system specific and that the whole chain from the PET system to the treatment PS must strictly be quality controlled when used in clinical practice.