The purpose of this study was to develop methodology to segment tumors on 18F-fluorodeoxyg- lucose (FDG) positron emission tomography (PET) images. Sixty-four metastatic bone tumors were included. Graph cut was used for tumor segmentation, with segmentation energy divided into unary and pairwise terms. Locally connected conditional random fields (LCRF) were proposed for the pairwise term. In LCRF, three-dimensional cubic window with length L was set for each voxel, and voxels within the window were considered for the pairwise term. Three other types of segmentation were applied: region-growing based on 35%, 40%, and 45% of the tumor maximum standardized uptake value (RG35, RG40, and RG45, respectively), SLIC superpixels (SS), and region-based active contour models (AC). To validate the tumor segmentation accuracy, dice similarity coefficients (DSC) were calculated between the result of each technique and manual segmentation. Differences in DSC were tested using the Wilcoxon signed-rank test. Mean DSCs for LCRF at L = 3, 5, 7, and 9 were 0.784, 0.801, 0.809, and 0.812, respectively. Mean DSCs for the other techniques were: RG35, 0.633; RG40, 0.675; RG45, 0.689; SS, 0.709; and AC, 0.758. The DSC differences between LCRF and other techniques were statistically significant (p < 0.05). Tumor segmentation was reliably performed with LCRF.
18F-fluorodeoxyglucose (FDG) positron emission tomography (PET) is widely used for the diagnosis and staging of malignant tumors. In addition, FDG-PET can be used to plan radiation therapy and assess treatment effectiveness. More than 1,700,000 clinical PET and positron emission tomography/computed tomography (PET/CT) studies were performed in the USA in 2011 [
Visual assessment and quantitative measurements based on region of interest, such as maximum standardized uptake value (SUVmax), have been clinically used to diagnose and stage malignant tumors using FDG-PET. FDG uptake has been studied thoroughly to allow its use as an imaging biomarker in oncology. One such application is tumor segmentation on FDG-PET images to quantify cancer aggressiveness. For example, it has been reported that metabolic tumor volume and total lesion glycolysis, based on functional tumor volume assessed from FDG- PET images, were more reliable predictors of tumor prognosis than SUVmax [
There are several problems for tumor segmentation on FDG-PET images. Although manual segmentation was often used because of its simplicity, it was time-consuming and susceptible to window level settings, and intra- and inter-observer variability [
To overcome these PET issues, this study proposes a new method of image segmentation based on graph cut [
This retrospective study was approved by the institutional review board of our institution. The acquisition of informed consent was waived by the review board.
Sixty-four bone tumors from 17 patients (11 men and 6 women; mean age, 61 years; range, 21 - 88 years) were included in this study. These bone tumors were clinically suspected to be metastases, and could be detected on FDG-PET images.
For each patient, whole body FDG-PET/CT images from the ear to the midthigh were acquired with a combined PET/CT scanner (Discovery 690; GE Healthcare, Waukesha, WI, USA). The PET component of the scanner can acquire 47 transaxial PET images in one bed position, with the following parameters: matrix size, 192 × 192 pixels; interslice spacing, 3.27 mm. The technical parameters for the CT component were: 16-detector row helical CT scanner; CT image matrix size, 512 × 512 pixels; interslice spacing, 3.27 mm. Patients received an intravenous injection of 160 - 320 MBq of FDG after at least 4 hours of fasting. About 60 minutes after the injection, low-dose CT was performed at a normal expiration position for attenuation correction of PET images. Then, an emission PET scan was performed at 2-min acquisition per bed position using a 3-D acquisition mode and time of flight method. Attenuation-corrected PET images were reconstructed with a 3-D ordered-subset expectation maximization reconstruction algorithm. In this study, only the PET images were analyzed.
To obtain the ground truth for the tumor segmentation, each bone tumor was manually delineated by a consensus between two board-certified radiologists (MN and AKK), who were also certified in diagnostic nuclear medicine. To reduce the computational cost, a 3-D bounding box surrounding the entire tumor was manually set by the radiologists. The radiologists evaluated the SUVmax of the tumor and the coordinates of the voxel where SUVmax was obtained (“the SUVmax voxel”). The FDG-PET images inside the bounding box, the SUVmax of the tumor, and the SUVmax voxel were used in the tumor segmentation. Pre- and post-processing were not utilized.
In graph cut, image segmentation is performed via energy minimization [
Generally, the segmentation energy of graph cut is divided into two terms, the unary term and the pairwise term. The energy E of the graph cut is represented as follows:
where k is a constant that controls the balance between the unary and pairwise terms; V is the set of all voxels within the PET images; l is a label (0 represents non-tumor and 1 represents tumor); U(v, l) is the unary energy term when assigning a label l to voxel v; N(v) is the set of neighborhood voxels around voxel v; and P(v, w) is the pairwise energy term, which is small when the similarity between voxel v and w is high. Using graph cut, the segmentation energy E can be minimized and labels can be assigned to the set of voxels V.
In grid conditional random fields (CRF), a 6- or 26-connected neighborhood is commonly used as N(v), and the segmentation energy represented by Equation (1) can be minimized by using graph cut. In this study, locally connected CRF (LCRF) was proposed as the pairwise term. Compared with the 6- or 26-connected neighborhood, LCRF incorporates local spatial information into the segmentation energy more widely, allowing more reliable tumor segmentation. In LCRF, a 3-D cubic window with length L was set for each voxel, and voxels within the window were considered for the pairwise term. The N(v) of LCRF was represented as the following equation:
where
where Iv and Iw are the SUV of voxel v and w, and dist(v, w) is the Euclidean distance between voxel v and w.
For the unary term, an adaptive threshold was used. First, the Otsu threshold [
where A is a parameter; and the mean(foreground_SUV) and mean(background_SUV) are the SUV means for the foreground and background voxels. The unary term U(v, l) is then defined as follows:
To compare the LCRF, three other types of segmentation methods were also applied: region-growing based on 35%, 40%, and 45% of the tumor’s SUVmax (RG35, RG40, and RG45, respectively), a region-based active contour model [
Region growing was performed for tumor segmentation using the SUVmax of the tumor and the SUVmax voxel as inputs. The seed point for region growing was the SUVmax voxel. The voxels were considered as candidates to be segmented if the SUV value of the voxel was greater than a lower threshold defined as a given percentage of SUVmax (in this study 35%, 40%, or 45%).
Several methods have been suggested as a region-based active contour (AC) model. In our study, the Selective Binary and Gaussian Filtering Regularized Level Set method was used, which can efficiently determine the contours even if the boundary of the tumor is weak or blurred [
Using SLIC superpixels (SS), sets of pixels can be divided into superpixels, which represent meaningful local structures [
In the current study, the following parameters were used for the segmentation methods:
1) RG35, RG40, and RG45 have no parameters.
2) SS has two parameters: superpixel compactness and the desired number of superpixels. Superpixel compactness ranged from 1 to 10, and the desired number of superpixels ranged from 1 to 20.
3) AC has two parameters: standard deviation of the Gaussian filter and the increase ratio of the level set function when solving the differential equation numerically. The standard deviation of the Gaussian filter ranged from 0.01 to 0.5. The increase ratio of the level set function ranged from 5 to 35.
4) In LCRF, there are three parameters: A of the unary term, L of the pairwise term, and k for controlling the unary term and pairwise term. The range of A was 0.3 - 0.7; that of L was 3 - 9; and that of k was 1 - 50.
To validate the accuracy of tumor segmentation, a Dice similarity coefficient (DSC) was calculated between the manual segmentation and the result of each technique. DSC is one of the most widelyused quantitative metrics to evaluate segmentation accuracy [
The DSCs of the segmentation techniques are shown in
Method | Mean of Dice similarity coefficient | Standard deviation of Dice similarity coefficient | p-values |
---|---|---|---|
RG35 | 0.633 | 0.253 | 8.32 × 10−8 |
RG40 | 0.675 | 0.225 | 1.14 × 10−6 |
RG45 | 0.689 | 0.214 | 4.45 × 10−5 |
SS | 0.709 | 0.171 | 4.09 × 10−8 |
AC | 0.758 | 0.177 | 0.0118 |
LCRF (L = 3) | 0.784 | 0.160 | 0.0151 |
LCRF (L = 5) | 0.801 | 0.135 | 0.118 |
LCRF (L = 7) | 0.809 | 0.139 | 0.407 |
LCRF (L = 9) | 0.812 | 0.136 | Not available |
Notes: p-values were results obtained from the Wilcoxon signed-rank test between LCRF at L = 9 and each of the other methods.
In this study, we proposed a new segmentation method for FDG-PET images based on graph cut. This method was applied to the clinical FDG-PET images of patients with metastatic bone tumors. The results of our study show that tumor segmentation was reliably performed using LCRF, and that the performance of LCRF was better than those of the other segmentation methods. In LCRF, the local spatial information was efficiently incorporated into the pairwise term of the segmentation energy, which led to accurate tumor segmentation.
Tumor segmentation and the determination of the tumor boundary on FDG-PET images are needed for quantitative evaluation of FDG-PET and guidance for localized cancer therapies (such as radiation therapy and interventional radiology procedures). Apart from manual segmentation, thresholding methods were the most widely used approach in clinical practice for delineating tumors on FDG-PET images. However, sensitivity to noise and inability to handle intensity variations made it difficult to perform tumor segmentation robustly in clinical cases using thresholding methods [
SS generates superpixels based on clustering of the local neighborhood [
The transition from the tumor to background can be gradual on FDG-PET images, which makes it difficult to segment a tumor accurately. AC, SS, and LCRF can handle this problem by using local or global information from PET images. However, when the SUV distribution of tumor was heterogeneous, the segmentation results using LCRF were more accurate than those of AC and SS, as shown in
The use of automatic or semi-automatic tumor segmentation has rarely been studied in bone tumor. To our knowledge, only Torigian et al. have evaluated the thresholding methods to segment spinal bone metastases on FDG-PET images [
There are several limitations to the current study. First, the numbers of patients and tumors included in this study were small or moderate. For further study, it is necessary to confirm the robustness and applicability of our segmentation method in a clinical setting by using a larger cohort of patients. Second, a widespread major problem for validating an image segmentation method is construction of the ground truth, and this problem was also found in the current study. Although it is highly desirable to validate the segmentation methods with pathology-based ground truth [
A new segmentation method based on graph cut was proposed for FDG-PET images. By using LCRF, local spatial information was efficiently incorporated into the segmentation energy of graph cut. Our results demonstrated that the results of LCRF surpassed those of conventional grid CRF, and that a tumor was more reliably segmented with LCRF than with other techniques.
MizuhoNishio,Atsushi K.Kono,KazuhiroKubo,HisanobuKoyama,TatsuyaNishii,KazuroSugimura,11, (2015) Tumor Segmentation on 18F FDG-PET Images Using Graph Cut and Local Spatial Information. Open Journal of Medical Imaging,05,174-181. doi: 10.4236/ojmi.2015.53022