Effect of different segmentation algorithms on metabolic tumor volume measured on 18F-FDG PET/CT of cervical primary squamous cell carcinoma.

BACKGROUND AND
PURPOSE: It is known that fluorine-18 fluorodeoxyglucose PET/computed tomography (CT) segmentation algorithms have an impact on the metabolic tumor volume (MTV). This leads to some uncertainties in PET/CT guidance of tumor radiotherapy. The aim of this study was to investigate the effect of segmentation algorithms on the PET/CT-based MTV and their correlations with the gross tumor volumes (GTVs) of cervical primary squamous cell carcinoma.
MATERIALS AND METHODS: Fifty-five patients with International Federation of Gynecology and Obstetrics stage Ia∼IIb and histologically proven cervical squamous cell carcinoma were enrolled. A fluorine-18 fluorodeoxyglucose PET/CT scan was performed before definitive surgery. GTV was measured on surgical specimens. MTVs were estimated on PET/CT scans using different segmentation algorithms, including a fixed percentage of the maximum standardized uptake value (20∼60% SUVmax) threshold and iterative adaptive algorithm. We divided all patients into four different groups according to the SUVmax within target volume. The comparisons of absolute values and percentage differences between MTVs by segmentation and GTV were performed in different SUVmax subgroups. The optimal threshold percentage was determined from MTV20%∼MTV60%, and was correlated with SUVmax. The correlation of MTViterative adaptive with GTV was also investigated.
RESULTS: MTV50% and MTV60% were similar to GTV in the SUVmax up to 5 (P>0.05). MTV30%∼MTV60% were similar to GTV (P>0.05) in the 5<SUVmax≤10 group. MTV20%∼MTV60% were similar to GTV (P>0.05) in the 10<SUVmax≤15 group. MTV20% and MTV30% were similar to GTV (P>0.05) in the SUVmax of at least 15 group. MTViterative adaptive was similar to GTV in both total and different SUVmax groups (P>0.05). Significant differences were observed among the fixed percentage method and the optimal threshold percentage was inversely correlated with SUVmax. The iterative adaptive segmentation algorithm led to the highest accuracy (6.66±50.83%). A significantly positive correlation was also observed between MTViterative adaptive and GTV (Pearson's correlation r=0.87, P<0.0001).
CONCLUSION: MTViterative adaptive is independent of SUVmax, more accurate, and correlated with GTV. Iterative adaptive algorithm segmentation may be more suitable than the fixed percentage threshold method to estimate the tumor volume of cervical primary squamous cell carcinoma.

Introduction

Cervical cancer is one of the most common malignant tumors in women. Radiation therapy is a commonly used treatment for Ia–IVa stage cervical cancer. Accurate delineation of the tumor radiotherapy target is critical to maximize the dosage to tumor tissue and to minimize the dosage to the surrounding normal tissues 1.

Metabolic tumor volume (MTV), measured on fluorine-18 fluorodeoxyglucose (18F-FDG) PET/computed tomography (CT) scan, represents the tumor biological target volume. It has the advantage of differentiating tumor tissue from surrounding normal tissues compared with conventional anatomic approaches 2–4. PET/CT-based MTV segmentation algorithms are usually based on visual inspection 5, absolute standardized uptake value (SUV) threshold 6, fixed SUVmax or fixed percentage of maximum standardized uptake value (SUVmax) threshold 7, and gradient method 8,9. At present, the fixed percentage of the SUVmax threshold algorithm is more popular, especially for target delineation of head and neck cancers, lung cancer, and cervical cancer 10.

In 2006, Sebastian et al. 11 published the iterative adaptive segmentation algorithm. The iterative adaptive algorithm has an advantage over fixed threshold methods in accurate delineation of the target volume according to the individual metabolic activity. This method is usually based on the SUVmax uptake within the volume and the threshold defined according to the background uptake within the adjacent normal tissue using a mathematical algorithm. This is then used to define the threshold and thus the 18F-FDG-avid volume of interest 12,13. However, the investigations of MTV delineation for cervical cancer using an iterative adaptive algorithm are still limited 14,15. This study compared the accuracy of MTV by the iterative adaptive algorithm (MTViterative adaptive) with that of the fixed percentage SUVmax threshold method using gross tumor volume (GTV) as the gold standard and investigated the correlation between them.

Materials and methods

Patients

The Institutional Research Ethic Committee of Shengjing Hospital approved the present study. All patients signed the informed consent for this study. Fifty-five patients from November 2012 to May 2015 with International Federation of Gynecology and Obstetrics stage Ia∼IIb were consecutively enrolled. 18F-FDG PET/CT scans were performed for staging 1 week before definitive surgery. All patients had histologically proved cervical squamous cell carcinoma. Patients were excluded if they had received chemotherapy or radiotherapy before surgery.

18F-FDG PET/CT imaging and analysis

All patients fasted for more than 6 h before 18F-FDG PET/CT scans were performed on them. Their serum glucose level was controlled to lower than 7 mmol/l. 18F-FDG (produced by MiniTrace II and TracerLab FX-FDG purity>99%; GE Healthcare, Waukesha, Wisconsin, USA) was injected at a dosage of 3.7 MBq/kg. Patients rested quietly for 60 min before PET/CT (Discovery PET/CT 690; GE Healthcare, USA) scanning. The wholebody CT and PET scans were all acquired with free breathing for attenuation correction and image fusion. First, low-dose non-enhanced CT images were scanned from the top of the skull to the mid-thigh, with bulb voltage 120 kV, Auto mA (30–210 mA; noise index 25), and slice thickness 3.27 mm, for attenuation correction of the PET images. Then, the PET data were acquired immediately after the CT scan using a three-dimensional acquisition mode at a speed of 1.5 min/bed (7∼8 beds in total) and a matrix size of 192×192. Attenuation-corrected PET images were reconstructed using an ordered-subsets expectation maximization iterative reconstruction algorithm, with 24 subsets and two iterations. Time-of-fight and point-spread function techniques were also used in the reconstruction.

Images were interpreted in a blinded manner by two experienced nuclear medicine physicians on an AW4.6 workstation (GE Healthcare, USA). All patients were divided into four groups according to the SUVmax within the target volume: SUVmax≤5 group, 5

MTV was defined as the volume of hypermetabolic tissue with an SUV greater than a defined threshold. The threshold was defined by a fixed percentage of SUVmax and an iterative adaptive algorithm. For the fixed percentage of the SUVmax method, an SUVmax threshold ranging from 20 to 60% was used in increments of 10% (MTV20%∼MTV60%). Volume Viewer (GE Healthcare, USA) was used to segment the fixed percentage of the SUVmax target. Volume computerized-assisted reporting (PET VCAR; GE Healthcare, USA) was used to segment MTV by the iterative adaptive algorithm (MTViterative adaptive) automatically. The PET and CT coregistration was first assessed once the images were loaded into the PET VCAR software. The primary PET gray scale and PET/CT fused images were then reviewed in the axial, sagittal, and coronal planes. Coregistration was assessed according to anatomical bony landmarks, for example, the base of the skull and vertebral pedicles. A boundary box, to autocontour and segment the region of interest, was placed over the image, reviewed, and adjusted to ensure that this three-dimensional cube contained the entire 18F-FDG PET-positive area and excluded the 18F-FDG PET-negative normal tissue. This process was repeated until each 18F-FDG PET/CT-positive region had been selected and optimized. These volumes were then segmented automatically using an iterative adaptive algorithm to detect the threshold level that separated the target volume from the background tissue by weighting the SUVmax and the SUVmean within the target volume with a weighting factor. This weighting factor was automatically set at 0.5 11,12. Percentage difference of segmentation MTVs and GTV was also calculated by (MTV−GTV)×100%/GTV and the optimal threshold percentage in the range 20∼60% for each patient was defined by the smallest percentage difference between segmentations results and the GTV.

Pathology

All patients underwent a radical hysterectomy and pelvic lymphadenectomy within a week after PET/CT scans. Once the fresh tumor specimen was collected after the surgery, the maximum vertical and horizontal tumor sizes were measured with a ruler. The specimens were subsequently fixed in 10% formalin for at least 24 h. The dimensions of the fixed tumor samples were also measured in the same way to determine the ratio of volume reduction because of fixation. Then, the fixed tumor specimens were sliced into consecutive tissue sections of 5 mm thick slices; the range of cutting included the entire tumor and surrounding normal tissues. Each slice was numbered and stained with hematoxylin and eosin and photographed using a digital camera. The tumor area (Ai) was outlined by an experienced pathologist with Photoshop 7.01 (Adobe Systems, San Jose, California, USA). The tumor area for each slide was measured and the GTV for each specimen was calculated as GTV=Σi Ai/R, where R was defined as the change ratio of the sample dimensions before and after fixation, respectively.

Comparisons of absolute values and percentage difference between MTVs by segmentation and GTV were performed in different SUVmax subgroups. The optimal threshold percentage was determined from MTV20%∼MTV60%, defined by the smallest percentage difference between segmentations results and the GTV. The optimal threshold percentage in different SUVmax groups was also compared and the relationship between them was analyzed. The correlation of MTV from iterative adaptive segmentation and GTV was also investigated.

Statistical analysis

Statistical analyses were carried out using SPSS 17.0 (SPSS Inc., Chicago, Illinois, USA). Continuous variables were presented as mean and SD. Independent t-tests were performed for the comparisons of MTVs by different tumor segmentation algorithms in different groups. Two-tailed Pearson’s correlations were used to assess the relationship between MTViterative adaptive and GTV. P value less than 0.05 was considered statistically significant (Fig. 1).

Fig. 1

A 54-year-old woman with moderately differentiated cervical squamous cell carcinoma and FIGO stage IIb1. MTVs of primary tumor were 10.13 cm3 (a), 6.35 cm3 (b), 4.04 cm3 (c), 2.87 cm3 (d), and 1.87 cm3 (e) by the 20∼60% SUVmax threshold, and 5.30 cm3 (f) by the iterative adaptive algorithm. GTV is 5.23 cm3. FIGO, International Federation of Gynecology and Obstetrics; GTV, gross tumor volume; MTVs, metabolic tumor volumes; SUVmax, maximum standardized uptake value.

Results

The mean age of 55 patients was 48.35±9.23 years, ranging from 26 to 68 years (Table 1). MTV20%, MTV30%, MTV40%, MTV50%, MTV60% and MTViterative adaptive were 23.01±16.99, 14.86±11.38, 10.21±8.76, 7.12±7.06, 4.86±5.26, and 12.56±10.60 cm3, respectively, and GTV was 12.35±10.10 cm3. Absolute MTV (cm3) and the difference percentage between MTVs of different segmentations and GTV in all subgroups are shown in Table 2 and Fig. 2. MTV50% and MTV60% were similar to GTV (P>0.05) in the SUVmax≤5 group. MTV30%∼MTV60% were similar to GTV (P>0.05) in the 5MTV20%∼MTV60% were similar to GTV (P>0.05) in the 10MTV20% and MTV30% were similar to GTV (P>0.05) in the SUVmax≥15 group. MTViterative adaptive were similar to GTV in both total and different SUVmax subgroups (P>0.05). Significant differences were observed among the different segmentation groups (Fig. 2). Twenty percent and 30% threshold consistently overestimated the MTV (121.4±111.5 and 41.9±66.31%), whereas 40, 50, and 60% thresholds led to underestimation of MTV (−8.34±42.47, −38.83±30.39, and −59.38±27.89%). The iterative adaptive segmentation algorithm yielded the highest accuracy (6.66±50.83%), although with a large SD 50.83%.

Table 1

Characteristics of all patients

Table 2

MTVs with different segmentation algorithms (mean±SD) and percentage difference in different SUVmax groups

Fig. 2

Relative percentage errors between GTV and MTVs of different segmentation methods. GTV, gross tumor volume; MTVs, metabolic tumor volumes.

Figure 3 shows the relationship between the optimal threshold percentage and SUVmax; the optimal percentage threshold was inversely correlated with SUVmax. The higher the SUVmax, the smaller the optimal threshold. In the SUVmax≤5 group, the optimal threshold percentage was 55±5.77%, whereas the optimal threshold percentage were 37.69±8.32, 35.88±7.12, and 30.48±8.65% in the 515 groups, respectively.

Fig. 3

Relationship between the optimal threshold percentage and SUVmax. SUVmax, maximum standardized uptake value.

Also, a significant positive correlation (r=0.87, P<0.0001) was observed between MTViterative adaptive and GTV (Fig. 4).

Fig. 4

Correlation between MTViterative adaptive and TGV (cm3). MTV, metabolic tumor volume; TGV, total glycolytic volume.

Discussion

It is known that accurate tumor target volume delineation is critical for the effectiveness of radiotherapy. As PET/CT shows the tumor metabolism activity, it is increasingly being accepted as an effective way for tumor biological target volume delineation 16–19. The aim of this study was to compare the MTVs from a PET/CT-based iterative adaptive algorithm and from the fixed percentage (20∼60%) of the SUVmax method to determine whether the MTVinterative adaptive algorithm is better estimated to GTV of primary cervical cancer.

Our results showed that the dynamic threshold percentage correlated inversely with SUVmax, similar to a previous report on esophageal tumor 20. However, the related trend between optimal percentage threshold and SUVmax was not obviously identical to that of Hyun et al. 21 study, in which the ideal threshold was defined when the maximum metabolic tumor length was exactly the same as the length of the known maximum pathologic tumor length. In our study, the optimal percentage threshold was relatively determined to be within 20∼60%, which was defined by the smallest percentage difference between MTVs by segmentations and the GTV; this was the main reason for the differences in the results. In addition, some authors applied a fixed percentage threshold to measure MTVs of cervical cancer and head and neck cancer 22–27. Kim et al. 22 compared MTV of 40% SUVmax threshold in 45 patients with invasive cervical cancer to their pathological and prognostic outcomes. The results showed that the patients with MTV larger than 20 cm3 had a significantly reduced disease-free survival compared with those with MTV less than 20 cm3 (P=0.029), and the MTVs were significantly different among the groups according to the status of lymph node metastasis, parametrial invasion, tumor differentiation, and International Federation of Gynecology and Obstetrics stage (P<0.05). In addition, a feasibility study by Ciernik and colleagues 25,26 evaluated the use of PET/CT in radiation therapy planning, and 50% SUVmax was found to be a reliable correlate to CT tumor volume. Showalter et al. 28 and Lee et al. 29 used the 40% SUVmax threshold to define the primary tumor diameter of early-stage cervical cancer and a significant positive correlation was found with the pathologic tumor diameter (r=0.757, P<0.0001). Also, a few other radiation therapy planning and treatment evaluation studies on cervical cancer used thresholds between 40 and 60% 22,30–31. Kidd et al. 30 found an MTV of 40% SUVmax threshold in cervical cancer decreased at weeks 2 and 4 after radiation therapy compared with the baseline (P<0.05). Yu et al. 31 evaluated the percentage SUVmax for non-small-cell lung cancer target volume delineation using pathologic GTV as the gold standard, they found that the optimal thresholds ranged from 20 to 42% (31±11%) SUVmax, and there was no certain optimal threshold. These findings suggest that different biological properties and metabolic levels of the primary tumor yield different percentages of the SUVmax threshold. Thus, the fixed percentage SUVmax threshold method is not suitable for different types of malignancy, tumor characteristics, and different metabolic levels 32, especially when SUVmax of at least 15. The volumes of the gross tumor are strongly dependent on the segmentation tools 13.

Our study confirmed that the MTV by fixed percentage threshold and the optimal threshold percentage are SUVmax dependent. Thus, fixed percentage SUVmax threshold was not the ideal method for MTV delineation, despite its simplicity of measurement for practical clinical use. It is obvious that an accuracy and maneuverable MTV segmentation algorithm is necessary for the planning of radiotherapy and accurate quantification of cervical cancer. This study adopted an iterative adaptive algorithm in the PET VCAR from GE Healthcare for automatic and accurate delineation of MTV in cervical cancer. Our results showed that MTViterative adaptive was similar to GTV in all SUVmax groups (t test, P>0.05), and yielded satisfactory results with the smallest bias (6.66%), although there was a large SD (50.83%) as MTVs by fixed percentage threshold were either largely overestimated or underestimated. Overall, MTViterative adaptive correlated strongly with GTV (r=0.87).

Some studies have reported a comparison of different methods for segmentation of 18F-FDG PET-positive tissue for target volume definition, but there are not many studies on iterative adaptive algorithm segmentation methods using SUVmax and SUVmean in tumor uptake. The iterative adaptive segmentation technique has some advantages compared with the traditional threshold algorithms. First, MTViterative adaptive is dependent on both the SUVmax and the SUVmean of the tumor uptake. Second, the accuracy of the iterations was assessed in the software by comparing the achieved and expected volumes 11,12. MTV from the iteration adaptive algorithm in our study showed a much better correlation with pathologic tumor volume compared with the fixed percentage SUVmax threshold method.

Yu et al. 33 used PET image X, Y, and Z axis maximum diameters compared with three corresponding directions of the maximum tumor pathological diameter to determine the optimal segmentation threshold. But, in this approach it is difficult to ensure that the PET image slide and gross specimens slice match well. The GTV measurement method in our study was more accurate because the entire tumor pathological volume estimation did not depend on the degree of pathological and corresponding PET imaging matching. In addition, we compared two volumes before and after fixing, and eliminated specimens for which the shrinkage rate was in excess of 5% of the first volume of the specimen to ensure the reliability of GTV measurement. There are some limitations in this study. First, our clinical samples are still limited. Second, although the change ratio of the sample dimensions before and after fixation was considered, manual measurement errors could not be avoided and we failed to overlap the volumes from PET imaging and histopathology because of the lack of reliable markers in pathologic sections. Third, the influence of tumor volume itself on the accuracy of the MTV estimation (i.e. the partial volume effect) was not analyzed. Thus, a large cohort study of MTV estimated by the iterative adaptive algorithm for cervical cancer radiotherapy target delineation or tumor growth/response is needed in the future to investigate its accuracy and repeatability.

Conclusion

MVTiterative adaptive is independent of the SUVmax, and is more accurate and correlated with GTV. Iterative adaptive algorithm segmentation may be more suitable than the fixed percentage threshold method to estimate the tumor volume of cervical primary squamous cell carcinoma.