Accurate FDG PET tumor segmentation using the peritumoral halo layer method: a study in patients with esophageal squamous cell carcinoma

F-18 fluorodeoxyglucose (FDG) PET/CT is a noninvasive modality for staging and localization of various malignant tumors that have high glucose metabolism. FDG PET/CT can provide a variety of metabolic parameters, such as maximum standardized uptake value (SUVmax), metabolic tumor volume (MTV) [1, 2], total lesion glycolysis (TLG) [3], and metabolic tumor length (ML) [4, 5]. Among these, SUVmax is the most commonly used FDG PET parameter because the measurement is easy and operator independent. Although SUVmax is a widely used parameter, it is a representation of the highest metabolic value of only one pixel within a metabolically active tumor. Thus, the total burden of the primary tumor cannot be evaluated by SUVmax.

MTV and TLG can provide volumetric metabolic information on malignant tumors. MTV is a measurement of the volume of a tumor with a high metabolism, while TLG is defined as the product of mean SUV and MTV [6]. A variety of methods have been used to measure and segment MTV. These include visual segmentation methods (SegVisual), where the region of interest is drawn manually [7, 8], fixed SUV threshold methods (SegSUV) [3, 6], fixed percentage of SUVmax threshold methods (Seg%) [3, 6], adaptive threshold methods (SegAdaptive) [6, 9], and gradient methods (SegGradient) [6, 10, 11]. Among these various segmentation approaches, SegSUV or Seg% is widely used to measure MTV because nearly all PET/CT workstations have an auto-segmentation program for fixed threshold methods. Accurate segmentation of MTV is also important for accurate measurement of TLG. Because TLG is a product of MTV and mean SUV (SUVmean), accurate MTV segmentation is essential for reliable TLG measurement. Furthermore, an accurate SUVmean estimate also requires accurate MTV segmentation. If segmented MTV is largely different from the true pathologic tumor volume, TLG will be inaccurate. ML is the length of the tumor on the PET image. The threshold determination for ML is similar to that for MTV [12].

SegAdaptive and SegGradient are known to be more accurate than SegSUV or Seg% [6, 10, 11]. The main principle of SegAdaptive is to adapt the percentage threshold of the phantom or tumor according to signal-to-background ratio [6, 9, 13]. Several optimal regression functions can be obtained by fitting various regression models [6, 9, 13, 14]. MTV using SegGradient is a method used to detect a large gradient change in radioactivity around the tumor. The large gradient change is located in the outmost peripheral portion of the tumor, and its location is used to define the tumor margin [11, 15, 16]. However, a specialized program and workstation, PET Edge® (MIM Software Inc., Cleveland, OH, USA), is needed to determine MTV using SegGradient.

Jun et al. [17] recently introduced a new method for MTV segmentation. They found a distinct layer between the tumor and background activity using a 10-step color scale with specific window level settings and named the distinct layer the “peritumoral halo layer (PHL).” Segmentation using the PHL method (SegPHL) was more reliable than MTV segmented by SegSUV or Seg%. Although SegPHL might be reliable, the previous study was performed in patients with a small and visually homogeneous type of tumor, namely papillary thyroid carcinoma [17]. Thus, further validation of large, heterogeneous tumors is needed to determine if SegPHL can be widely applied in clinical settings.

Our aim was to compare variously segmented MLs (i.e., ML segmented by SUV 2.5 (ML_2.5), 40% of SUVmax (ML_40%), and PHL (ML_PHL)) of esophageal squamous cell carcinoma (ESCC) with pathological tumor length (PL) and to identify whether SegPHL could be used to reliably define tumor margin in ESCC, which can be large and/or heterogeneous.

The major finding of the current study is that SegPHL is more accurate and reliable for defining the tumor margin of ESCCs than are fixed threshold methods. In a previous study, Jun et al. [17] reported that SegPHL was more reliable than fixed threshold methods for delineation of the margins of thyroid cancers. However, the thyroid cancers examined in that study were very small and visually homogeneous. They reported that PHL abruptly increased in layer thickness and was located between the tumor and the background. In the present study, we demonstrated that SegPHL was also accurate in ESCCs, which can be large and/or visually heterogeneous. ML_PHL was highly correlated with PL (Fig. 5), the bias and standard deviation of ML_PHL were small on the Bland-Altman plot (Fig. 5), and there were no serious proportional errors visible on the plot. The absolute difference between ML_PHL and PL was smaller than that of ML_2.5 and PL (Fig. 6). ML_PHL was less affected by SUVmax of the ESCC than were ML_2.5 and ML_40%. The PHL tumor threshold (% of SUVmax) showed an inverse correlation with SUVmax (Fig. 7). The characteristics of the PHL were similar to those reported in a previous study [17]; PHL also showed an abrupt increase in layer thickness and was located between the tumor activity and the background. Furthermore, PHL tumor threshold changed according to SUVmax of the ESCC (Fig. 7), consistent with a previous PHL study [17].

The most widely used segmentation methods are SegSUV and Seg%. However, accurate tumor segmentation with a fixed threshold is impossible with these methods because the optimal tumor threshold changes according to SUVmax or signal-to-background ratio [13, 20, 21]. The volume of a tumor with faint FDG uptake is highly overestimated by Seg%. In addition, MTV using SegSUV underestimates the MTV of a tumor with low metabolic activity and overestimates that of a tumor with high metabolic activity [6]. These shortcomings of Seg% and SegSUV were also found in our present study. The difference between ML_2.5 and PL was larger when SUVmax was higher, and the difference between ML_40% and PL was larger when SUVmax was lower. However, ML_PHL was less affected by SUVmax. The results of our study suggest that the PHL method could accurately segment FDG avid tumor regardless of SUVmax.

SegVisual has been reported to be more accurate than the fixed threshold method in some cases [7]. Although accurate tumor segmentation may be possible with visual tumor delineation, the SegVisual has several limitations. First, the reproducibility of SegVisual is low [22]. Second, visual tumor perception can be altered by the PET window level setting and the color scale [23‐25]. Third, the process is quite slow, because no auto-segmentation program is used.

The SegPHL could be considered as a SegVisual because it involves the visual location of the background and the PHL. However, there are several major differences between SegPHL and SegVisual. First, inter-observer agreement for the SegPHL is very good (Table 3), unlike with SegVisual. This good agreement may be due to the use of specific PHL image settings (i.e., PET window level: top = SUVmax, bottom = 0; color scale = 10-step color scale). Among our study population of 121 cases, PHL tumor threshold was concordant in 111 (91.7%). The second difference between SegPHL and SegVisual is that SegPHL uses commercial auto-segmentation programs to measure MTV_PHL with a known, accurate PHL tumor threshold (i.e., % of SUVmax). Use of an auto-segmentation program allows more rapid measurement of MTV than is possible with SegVisual.

SegPHL may also be similar to SegGradient. The unique characteristics of PHL are location between the tumor and the background and abrupt increase in layer thickness compared to the inner layer without significant distortion of tumor shape [17]. In the gradient method, the tumor margin is considered the point at which a large gradient change occurs in the PET/CT image [11]. A previous study suggested that PHL was very similar to the large gradient change of SegGradient [17]. We also hypothesized that the location of the PHL would likely be very similar to that of the large gradient change of SegGradient. Furthermore, like SegGradient [11], SegPHL was more accurate than fixed threshold methods. The main differences between SegPHL and SegGradient are as follows. First, PHL tumor threshold is determined by visual inspection of the tumor using PHL image settings. In contrast, SegGradient is determined using a special workstation. Second, an accurate percentage threshold can be determined by SegPHL (i.e., % of SUVmax). However, there is no way to determine an accurate % threshold of SegGradient. Third, although SegPHL is similar to SegGradient, the exact point at which there is a large gradient change cannot be determined using SegPHL. In our study, we defined PHL tumor threshold as the starting point of the PHL (i.e., the innermost point of the PHL), as in Jun et al. [17]. However, the large gradient change of SegGradient could be located in the outermost point or in the middle of the PHL. Another consideration is that the large gradient change of SegGradient might not be represented by a single % of SUVmax (i.e., the SUV of the large gradient change may differ in different parts of the tumor).

SegAdaptive is based on calibrated curves, where accurate % threshold changes as signal-to-background ratio changes [6, 13]. In this method, as the signal-to-background ratio increases, the accurate % threshold decreases [13]. Clinicians can choose an optimal regression equation to determine the accurate % threshold. Based on a phantom study, SegAdaptive is superior to fixed threshold methods [13, 26]. In our present study, PHL tumor threshold showed a strong inverse correlation with SUVmax of ESCC (σ = − 0.923, p < 0.0001; Fig. 7a), as reported for SegAdaptive in the phantom study. Similarly, Jun et al. [17] reported that PHL tumor threshold was inversely correlated with SUVmax of thyroid cancer.

Several studies have compared ML, MTV, PL, and pathologic tumor volume [12, 21, 27, 28]. The results of these reports suggest that there is no single optimal SUV or % of SUVmax threshold that can be used to define an accurate tumor margin [6]. In our present study, PHL tumor threshold (%) showed a significant inverse correlation with SUVmax of ESCC (σ = − 0.923, p < 0.0001; Fig. 7a), and SUV of PHL tumor threshold had a significant positive correlation with SUVmax of ESCC (σ = 0.891, p < 0.0001; Fig. 7b). Our results are consistent with previous studies that compared ML, MTV, PL, and pathologic tumor volume [12, 21, 27] and with a previous PHL study of thyroid cancer [17]. The number of cases where ML was compared with PL was similar between our present study and other previous studies (Zhong et al. [12]: n = 37 in esophageal cancer; Borakati et al. [28]: n = 21 in esophageal cancer).

An accurate tumor segmentation method should satisfy the following conditions. First, the segmented ML has to have a high degree of correlation with PL. Second, the bias and standard deviation should be small on Bland-Altman plots. Third, no serious proportional errors should be found in Bland-Altman analysis. Fourth, the segmented ML should not be highly affected by SUVmax of ESCC. Fifth, the segmented ML should correlate strongly with EL. Among ML_PHL, ML_2.5, and ML_40%, only ML_PHL satisfied these conditions. Although ESCCs in our present study were definitely larger and more visually heterogeneous than the thyroid cancers examined in a previous PHL study [17], SegPHL was the most accurate segmentation method among SegPHL, SegSUV, and Seg%. Therefore, we conclude that SegPHL is far more accurate than fixed threshold methods.

Our study had several limitations. First, the study design was a retrospective analysis. Thus, there might have been operator-dependent bias in the measurement of EL and PL. Second, comparison of MTV and pathologic tumor volume was impossible because a vertical incision along the esophagectomy specimen had to be performed to evaluate pathologic depth of invasion. The vertical incision process changed the ellipsoid tumor shape of the ESCC into a planar shape. Because a change in tumor shape could introduce serious bias, we only assessed the correlation between ML and PL. Third, PL was measured after formalin fixation, which could shrink the esophagectomy specimen.

In conclusion, SegPHL was more accurate than fixed threshold methods compared with PL. PHL tumor threshold was adjusted according to SUVmax of ESCC. We conclude that SegPHL might be used to accurately define the tumor margins of ESCCs.