A methodological investigation of healthy tissue, hepatocellular carcinoma, and other lesions with dynamic 68Ga-FAPI-04 PET/CT imaging

Background The study aimed to establish a 68Ga-FAPI-04 kinetic model in hepatic lesions, to determine the potential role of kinetic parameters in the differentiation of hepatocellular carcinoma (HCC) from non-HCC lesions. Material and methods Time activity curves (TACs) were extracted from seven HCC lesions and five non-HCC lesions obtained from 68Ga-FAPI-04 dynamic positron emission tomography (PET) scans of eight patients. Three kinetic models were applied to the TACs, using image-derived hepatic artery and/or portal vein as input functions. The maximum standardized uptake value (SUVmax) was taken for the lesions, the hepatic artery, and for the portal veins—the mean SUV for all healthy regions. The optimum model was chosen after applying the Schwartz information criteria to the TACs, differences in model parameters between HCC, non-HCC lesions, and healthy tissue were evaluated with the ANOVA test. Results A reversible two-tissue compartment model using both the arterial as well as venous input function was most preferred and showed significant differences in the kinetic parameters VND, VT, and BPND between HCC, non-HCC lesions, and healthy regions (p < 0.01). Conclusion Several model parameters derived from a two-tissue compartment kinetic model with two image-derived input function from vein and aorta and using SUVmax allow a differentiation between HCC and non-HCC lesions, obtained from dynamically performed PET scans using FAPI.


Background
Hepatocellular carcinoma (HCC), the most common primary liver cancer. HCC remains an important global clinical challenge due to its hidden onset. It is a highly heterogeneous cancer [1]. The forms of heterogeneity are seen when comparing tumors between patients (interpatient heterogeneity), between different tumor nodules within the same patient (intertumoral heterogeneity), and between different regions of the same nodule (intratumoral heterogeneity) [2]. Although imaging modalities, including ultrasonography, CT, and MRI are valuable in hepatic lesion characterization, they still have limitations in distinguishing the functional variables or the differentiation of malignant lesions [1]. There were shreds of evidence that limited sensitivity using 2-deoxy-2-[18F]fluoro-D-glucose (FDG) PET in detecting hepatocellular carcinoma with the falsenegative rate approaches 40-50% [3]. Thus, the limited diagnostic efficacy of current imaging strategies remains a major challenge in the accurate evaluation of hepatic lesions.
Fibroblast activation protein (FAP) is overexpressed in cancer-associated fibroblasts (CAFs) in several tumor entities, especially in breast, colon, and pancreatic carcinomas [4]. It plays a variety of cancer-promoting roles and may predict a poor prognosis [5]. FAP-specific inhibitors (FAPIs) have already been developed as anti-cancer drugs. Recently, its quinoline-based derivatives have been designed into radiopharmaceutical agents [6,7]. FAPI PET demonstrated promising preclinical and clinical results. A few first-in-human studies have found superior tumor-imaging potential of FAPI over FDG PET in different tumor entities [5,6,8,9]. Due to low background in muscle, bloodpool and liver, the tumor-to-background contrast ratios are significantly higher. This may indicate that FAPI PET is more likely to detect tumors in the liver when FDG PET faces limitations [10].
Several investigators suggested dynamic PET with kinetic modeling might have potential value in the differential diagnosis other than static quantitative parameters like SUV [11][12][13]. Since the liver has blood supply from both the hepatic artery and portal vein, which should be both considered as the input function, so an accurate model is crucial for the following analysis. Few different kinetic models of FDG and 11 C-acetate in the liver have been established. In our previous study, we compared several FDG dynamic models and introduced a simple two-tissue model using the portal vein solely to differentiate HCC from healthy liver tissues [14] accurately.
To our knowledge, the significance of FAPI PET kinetics in the liver has not been reported yet. Since the metabolic process at the cellular level is different between FAPI (binding to FAP and internalization) and FDG (entering glycolysis and accumulation), our previous liver FDG PET kinetic model [14] might not be able to fit FAPI PET kinetics. The main objective of this study, therefore, is to establish a kinetic model for 68 Ga-FAPI-04 kinetics in distinguished hepatic lesions which were histologically determined, as well as to evaluate if the model-derived parameters allow a differentiation between HCC and non-HCC lesions non-invasively.

Patient characteristics
The study was approved by the Ethics Committee at Peking Union Medical College Hospital. All patients signed informed consent. Eight male patients with 12 available liver lesions (age range, 47-70 years) were recruited in this study. The pathology of 7 lesions was confirmed by surgical resection and the pathology of 5 lesions was by needle biopsy. Among these patients, four patients had been histologically confirmed as hepatocellular carcinoma (HCC), two patients had intrahepatic cholangiocarcinoma, one patient had liver metastasis of gastric cardia adenocarcinoma and one had inflammatory granulomatous.
PET/CT scan PET/CT scans were conducted on a PoleStar m660 PET/CT scanner (SinoUnion Healthcare, Beijing, China) at Peking Union Medical College Hospital (PUMCH) [15] for all patients. CT transmission scans (120 kV, 260 effective mA) were performed first for attenuation correction and image fusion. Then, 174-259 MBq 68 Ga-FAPI-04 was administered intravenously and the dynamic PET was performed over the liver region simultaneously. Each PET scan lasted 60 min. Dynamic PET images were reconstructed using a manufacturer (SinoUnion Healthcare, Beijing, China) who provided a stand-alone advanced research workstation with standard ordered subset expectation maximization (OSEM) algorithm with 2 iterations and 10 subsets. The 120-frame reconstruction protocol consisted of 60 frames of 5 s, 10 frames of 30 s, and 50 frames of 60 s.

Image analysis
Delineation of volumes of interest was done on the Hermes Hybrid Viewer tool (Hermes Medical Solutions AB, Stockholm, Sweden). The volumes of interest (VOIs) were drawn manually over all visible lesions and healthy regions within an area far away from any lesion in the liver on the CT image of each patient. For the image-derived input functions (IDIF), a VOI was drawn within the hepatic artery (here denoted as A) and one in the portal vein (V). Then VOIs were copied to the dynamic PET images. The corresponding concentration time-activity curves (TACs) were extracted. In case of healthy regions, the mean standardized uptake value (SUV) was taken from each region. In case of the lesions and the IDIFs, the voxel with the maximum SUV (SUV max ) was taken, since these values are least affected by partial volume and motion effects [16]. Furthermore, the according volume sizes and the uptake of the last 5 min were exported. A representative PET/CT scan is shown in Fig. 1.

Kinetic models
To describe tracer kinetic processes in a certain organ, a kinetic model is applied according to TAC. This usually is a set of differential equations that describe all underlying pharmacokinetic processes using so-called compartments, which are solved based on an input function and fitted to the TACs. As a result, rate constants (denoted with k) are obtained reflecting the biochemical exchange processes of the tracer. The kinetic behavior of tracers in the liver is usually described with a two-tissue compartment model using two input functions [17], since the liver tissue is supplied by both, an arterial and a venous blood input. As described above, IDIFs were used from the hepatic artery and the portal vein, both fitted with a tri-exponential function starting from the peak maximum and with a linear increase before the maximum. To find the optimum model, three different kinetic models were applied to all TACs (Fig. 2). Two reversible two-tissue compartment models were applied, as usually applied on TAC obtained from FDG PET scans: one using a single input function from the hepatic artery, denoted as "model A," and a second one also with a single input function from the portal vein, denoted as "model V." Additionally, a third, also with two-tissue compartmental "model AV" using both input functions from artery and vein was applied, according to the formulas: with C 1 (t) and C 2 (t) as the hepatic compartment, A(t) as the artery IDIF, and V(t) as the venous IDIF. Since A(t) and V(t) are both contributing to the fraction of blood volume by having a proportion of K a /(K a + K v ) and K v /(K a + K v ) to v B (see Eq. 3 below), each amount was assessed from their inflow rate constants for the measured compartment C measured (t): with v B as the fraction of the measured volume occupied by blood and K a and K v as the influx rate constant of the aorta and the vein, respectively.
With the rate constants as fit parameters, all model fits were performed according to the least-squares method and optimized with Levenberg-Marquardt algorithm, implemented to a Java program. Errors of the fit parameters were estimated by calculating the covariance matrix. The residual sum of squares was calculated for each TAC, as well as the average and standard deviation of all rate constants; furthermore, the parameters V T = K 1 /k 2 (1 + k 3 /k 4 ), BP ND = k 3 /k 4 , and V ND = K 1 /k 2 were calculated, with K 1 as either K a , K v or (K a + K v ) depending on the model. Since to our knowledge no initial parameters are available for FAPI model parameters, every TAC was first fitted with a one-tissue model to obtain values for K 1 , k 2 , and v B . For the two-tissue models, these parameters were taken as initial values and a second fit was performed to obtain k 3 and k 4 ; note that the latter could become zero thus resulting in a one-tissue model for the corresponding TAC.

Statistical analysis
In order to find the optimum model, the Schwartz criterion (SC) [18] was applied to each TAC. In detail, for each model, the number of those TACs was counted, which were preferred according to the SC and subsequently expressed as a percentage of all TACs. A Kolmogorov-Smirnov test was performed to check for normal distribution. The ANOVA test was used to find differences between HCC, non-HCC and healthy regions, significant differences between two groups were estimated with the unpaired Student's t test, or the Wilcoxon-Mann-Whitney test if not normally distributed. Due to the small sample size, a p value of less than 0.01 was classified as significant.

Lesion characteristics
Eight patients with 12 available liver lesions were recruited in this study. The 12 available lesions were separated according to histological results into a group of 7 HCC lesions (group HCC) and a group of 4 other lesions (2 ICC and 2 gastric metastases lesions). One lesion was an inflammation, having a TAC and uptake very similar to healthy tissue, it therefore was treated separately. Lesion characteristics are summarized in Table 1. All values showed a possible normal distribution, except for the parameters SUV, V T , and V ND from the healthy regions.

Model selection
For all TACs, model AV was preferred in 70% of all cases, in detail, in 58% of all lesions, in 71% of only HCC lesions, and in 88% of all healthy regions, see Fig. 3. This distribution did not change after excluding the inflammation lesion. All relevant mean parameters derived from model AV are listed in Table 2 for all groups separately. With only one exception, k 3 and k 4 were zero in the case of all healthy regions.

Comparison between HCC and non-HCC
No significant differences were found in the SUV uptake according to the ANOVA test between all different groups HCC, non-HCC, inflammation and healthy (p = 0.06), nor between only HCC and non-HCC group (8.1 versus 13.1, p = 0.08), see Fig. 4. However, comparing only the SUV uptake from all lesions together (HCC and non-HCC) versus the healthy regions, a significant difference was found (9.9 versus 0.8, p = 0.0002 according to both, Student's t test and Wilcoxon-Mann-Whitney test). Including the inflammation lesion to the lesion group, the difference was slightly less significant with p = 0.0001 (9.9 versus 0.8). Regarding volume sizes, no significant differences were found between HCC and non-HCC groups without the inflammation (6 cm 3 versus 23 cm 3 , p = 0.09).
Note that for all applied models, the differences in k 3 and k 4 were always significant between HCC and non-HCC lesions compared to healthy regions, since k 3 and k 4 were equal or close to zero for all models and all healthy regions. For models A and V, there were no significant differences found between HCC and non-HCC lesions in case of all rate constants and the macro-parameters V T (p = 0.08 and p = 0.10, for model A and Table 1 Standardized uptake value (SUV) and volume size of all investigated lesions, including the preferred model for each lesion. For the healthy region, the average overall regions is presented for SUV mean , size, and preferred model V, respectively) and V ND (p = 0.73 and p = 0.57). In case of BP ND , a significant difference was found only for model A: 2.9 for HCC and 6.4 for non-HCC, p = 0.0006. With model AV, significant differences according to the ANOVA test were found between all groups in case of the parameters V T (p < 0.0001), V ND (p < 0.0001), and BP ND (p< 0.0001). Including the inflammation, lesion did not change this finding (all according p < 0.001). Comparing all groups among each other without healthy regions (which had k 3 and k 4 of almost or equal zero), a significant difference was also found in BP ND (p = 0.003). A direct comparison of HCC and non-HCC lesions (without inflammation) showed significant differences in V T (3.7 versus 7.6, p = 0.002), V ND (0.9 versus 1.2, p = 0.29), and BP ND (3.2 versus 5.6, p = 0.006).

Discussion
Noninvasive assessment of liver mass is crucial in differentiating hepatic malignancy. Many cases are already advanced at the time of diagnosis since the occult onset and lack of effective biomarkers for hepatic malignancy [19,20]. Our findings showed preliminary results that using dynamic FAPI PET with kinetic modeling could identify  However, there were no reliable corrections for partial volume or motion effects available; therefore, the values in particular for K a , K v , and k 2 may not be taken as real values with biological meaning. Although these effects are minor in the case of the aorta [21,22], this is certainly not the case for smaller blood vessels like the portal vein. However, to reduce these effects, SUV max was taken for all IDIFs [16], and also for all lesions due to their partially very small sizes.
Although only minor knowledge is known of FAPI kinetics, it seemed obvious to choose two-tissue compartmental models, since FAPI apparently was trapped in the organ and the lesions. We assume that in case of all lesions, a one-tissue compartmental model (i.e., k 3.4 = 0) was not sufficient, because the fits resulted in values of k 3,4 > 0 for all lesions are values of k 3,4 ≈ 0 for all healthy regions.
Not surprisingly, the preferred model was a model taking venous and arterial input into account, since liver tissue in general is supplied by both [17]. Using this model, significant differences could be found between the four investigated groups, HCC, non-HCC, and healthy regions in case of all rate constants, the binding potential BP ND , the total distribution volume V T , and distribution volume of non-displaceable uptake V ND . Note that the results were not changed significantly after including the inflammation lesion into this analysis. These results therefore suggest that there are kinetic differences in FAPI kinetic between HCC and non-HCC lesions. Also not surprising, k 3 and k 4 were zero for almost all healthy regions, meaning that a reversible one-tissue model described healthy tissue best.
The inflammation lesion was treated separately since its SUV uptake as well as its curve shape was very similar to healthy tissue having also similar model parameters. A wrong delineation aside from the real lesion due to motion effects can be excluded: a shifting of PET data due to motion effects would most probably affect the whole organ; thus, similar curve shapes due to wrong delineations would have been observed also in other, even smaller lesions, which was not the case. Thus, if the delineation was correct, we conclude that FAPI tracer kinetics might be similar to healthy regions in case of inflammation. Fig. 4 For HCC lesions, non-HCC lesions, one inflammation, and healthy regions, the SUV uptake for the last 5 min (left) as well as the parameter V T derived from model AV (right) are displayed. Differences between HCC and non-HCC lesions were not significant in case of the SUV uptake (averaged 61 % difference, p = 0.08), and significant in case of V T (averaged 106% difference, p = 0.002) A major limitation of this study certainly is the small cohort size. However, this was a first proof-of-concept study showing that IDIFs (with a simple compensation method) are sufficient to derive relevant model parameters from describing FAPI kinetics in liver lesions. Also, the lesions are very heterogeneous, in particular, the one inflammatory lesion was difficult to handle. Therefore, it was treated separately and the analysis was done twice, with and without including the inflammation. Moreover, due to a missing reliable motion and partial volume corrections, values like K 1 and k 2 cannot be taken as biologically relevant; however, the macro parameters V ND , BP ND , and mainly V T show significant differences between all groups. Further studies, with a reliable integrated correction method and including a comparison to pathological results, will be conducted.

Conclusion
In the present study, we investigated kinetic models for FAPI PET in liver lesions, showing that the consideration of the maximum SUV values for artery and venous image-derived input function is suitable to at least distinguish between healthy regions, HCC lesions, and non-HCC lesions in model-derived macro parameters.