18F-fluoro-2-deoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT) is widely used for the initial diagnosis, restaging, and treatment response evaluation of the many kinds of tumors [1]. Among the multiple parameters that may be obtained from the FDG PET/CT, the standardized uptake value (SUV) is generally measured and accepted as an effective index [2]. In several previous studies, it was reported that tumor maximum SUV (SUVmax) is related with the prognosis of cancers [3,4,5]. However, SUV has some limitations. The SUV measurement can be affected by many factors including time, blood glucose concentration, and partial volume effects [2]. SUVmax does not reflect the metabolic activity of the entire tumor, representing only the maximum SUV in a voxel contained within a tumor region-of-interest [6]. Also, in some tumors, the SUVmax is not correlated with the prognosis [7, 8]. Due to the limitations of the SUV, it is difficult to use only SUVmax for the prediction of tumor prognosis, and other significant PET indexes are needed. Parameters including metabolic tumor volume (MTV) and total lesion glycolysis (TLG) began to emerge compensating the role of SUV [1]. It was reported that MTV (one of the PET parameters) is related to the prognosis of various cancers [1, 3, 9, 10].
The definition of MTV, which is related to the distribution of metabolic activity, is the volume of hypermetabolic tissue that has metabolic activity exceeding a defined threshold [11]. In order to accurately measure MTV for cancer prognosis, various PET tumor segmentation methods have been attempted [12]. These various conventional methods include the absolute SUV threshold method (e.g., SUV 2.0), fixed percentage SUVmax threshold method (e.g., 30% SUVmax), and signal-to-background method [9, 12]. However, the gross MTV is measured differently according to the various segmentation methods [12]. There is no standard method for measuring MTV. Therefore, among the various methods that are currently in use, the one that can best serve as a reference method remains controversial. [12].
Multi-level Otsu methods have been applied in several other application areas including in segmentation problems related to CT images. In the field of PET imaging, a variation of the basic Otsu method has been introduced as a solution to the PET segmentation problem [13]. However, in our literature search, we did not find any prior work related to the use of multi-level Otsu threshold technique applied to PET. We have applied this multi-level Otsu method to PET segmentation (MO-PET), as previously reported [14, 15]. It was demonstrated that MO-PET segmentation method is relatively accurate, stable, and consistent across a range of lesion sizes and PET lesion-to-background ratios representative of clinical tumor lesions [14, 15]. This MO-PET algorithm and method is summarized below and detailed in this reference [16] (https://www.google.com/patents/WO2016160538A1?cl=en).
Multi-level Otsu method, based on a more commonly known image threshold method known as Otsu’s method [17], is a simple and very effective clustering-based approach to convert a gray-level image to a binary image. The original Otsu method assumes that the image contains two classes of pixels (e.g., foreground and background) then calculates the optimum threshold that separates the two classes of pixels. The optimum threshold is computed such that the intra-class variance between the two classes of foreground and background pixels is minimal, which also corresponds to maximizing the inter-class variance between the two classes of the pixels. Multi-level Otsu method represents an extension of the same basic idea, i.e., minimization of the intra-class variance (which in turn results in maximization of inter-class variance) to images that contain clusters of pixel populations representing different structures that can therefore be classified at multiple threshold levels. Mathematically, MO-PET algorithm expands the original equation in the Otsu method for two pixel group classifications into an equation for classifying into an arbitrary number of classes. Thus, given the probability of occurrence of a pixel value i given by P
i
, the algorithm calculates the mean pixel level (μ) of the image and the inter-class variance (σ):
Mean level
$$ {\mu}_1\left({T}_1\right)=\sum_{i=1}^{T_1}i\frac{P_i}{P_1\left({T}_1\right)} $$
$$ {\mu}_2\left({T}_2\right)=\sum_{i={T}_1}^{T_2}i\frac{P_i}{P_2\left({T}_2\right)} $$
$$ {\mu}_K\left({T}_{K-1}\right)=\sum_{i={T}_{K-1}}^Li\frac{P_i}{P_K\left({T}_{K-1}\right)} $$
Variance
$$ {\sigma}_1^2\left({T}_1\right)=\sum_{i=1}^{T_1}{\left(i-{\mu}_1\left({T}_1\right)\right)}^2\frac{P_i}{P_1\left({T}_1\right)} $$
$$ {\sigma}_2^2\left({T}_2\right)=\sum_{i={T}_1}^{T_2}{\left(i-{\mu}_2\left({T}_2\right)\right)}^2\frac{P_i}{P_2\left({T}_2\right)} $$
$$ {\sigma}_K^2\left({T}_{K-1}\right)=\sum_{i={T}_{K-1}}^L{\left(i-{\mu}_K\left({T}_{K-1}\right)\right)}^2\frac{P_i}{P_K\left({T}_{K-1}\right)}, $$
where i is an individual SUV value (within the SUV range), L is the maximum SUV level in a given image, and T1, T2, … TK-1 are multiple threshold levels that can potentially be computed in a given image based on the distribution of the SUV within the image or a region-of-interest. Multiple threshold level values are determined by exhaustively searching through all sets of threshold levels for the given number of classes (K) in to which the image needs to be divided to find the combination that gives the minimum within class variance (or maximum inter-class variance). Thus, ultimately, the algorithm generates K classes and K-1 thresholds for a given image.
In this research, MO-PET, an automatic algorithm requiring very minimal user-input, was used for measuring the MTV of soft tissue sarcoma. MTVs measured with MO-PET and other conventional methods were compared in order to evaluate the usefulness and robustness of MO-PET.