Introduction

Vascular invasion, along with tumor size and the number of nodules, is a well-known major risk factor of early relapse of hepatocellular carcinoma (HCC) after surgical treatment [1, 2]. Macrovascular invasion can be detected by visualizing a tumor in the vein, whereas microvascular invasion (MVI) can only be microscopically detected. Furthermore, "macrotrabecular-massive" (MTM), a histological subtype of HCC that has been identified by recent studies, statistically correlated with MVI, is an independent predictor of early and overall relapse [3,4,5]. The role of preoperative imaging in assessing vascular invasion is still limited except when hepatic venous or portal venous tumor thrombus can be diagnosed [6, 7]. Some interesting studies have shown that suspicious qualitative imaging features, such as non-smooth tumor margins, internal arteries, mosaic architecture, irregular arterial rim-like enhancement, peritumoral arterial phase enhancement, and peritumoral hepatobiliary phase hypointensity, could predict MVI in HCC [8,9,10,11,12,13,14]. In order to improve imaging diagnosis of aggressive HCCs, recent studies have used radiomic analyses based on supervised machine-learning, in spite of major remaining challenges such as optimize feature extraction in radiological images [15,16,17,18].

Another approach to the prediction of preoperative vascular invasion is the use of quantitative imaging techniques. The results of perfusion computed tomography (CT) and dual-energy CT (DECT) data have been shown to be highly promising in the noninvasive assessment of tumor characteristics [19, 20]. Indeed, quantitative CT perfusion parameters, which are used to quantify tumor vascularity and angiogenesis in patients with HCC, reflect tumor aggressiveness and thus, could also help determine the prognosis [19, 21]. Furthermore, DECT-based iodine maps have been considered to be a potential perfusion biomarker of HCC [22]. Iodine-specific maps can reveal the presence or absence of local iodine in tissues and in the first case allow for its quantification [22]. It has been suggested that tumor response to anti-angiogenic therapy in patients with HCC may be predicted through the quantification of tumor iodine uptake as a biomarker of tumor vascularity [23]. Recent studies have also shown correlations between iodine density and some perfusion parameters in patients with HCC [24,25,26]. Therefore, the combined analysis of perfusion CT and DECT parameters assessing HCC tumor vascularity, should provide quantitative data in addition to the qualitative visual interpretation.

The study aims to evaluate perfusion CT and DECT quantitative parameters for predicting MVI of HCC prior to surgery.

Materials and methods

Patients

Patients were prospectively and consecutively enrolled in this single-center study between December 2017 and March 2020, after providing written informed consent. The study was approved by the local institution review board. The study was registered with this provided number: ClinicalTrials.gov: NCT03754192. Inclusion criteria were (a) a diagnosis of HCC (≥ 20 mm) according to American Association for the Study of Liver Disease guidelines, (b) curative intent hepatic resection recommended by a multidisciplinary team, and (c) preoperative CT performed within 2 months prior to surgery in our radiology unit. Exclusion criteria were (a) previous locoregional or systemic HCC treatment, and (b) a contraindication to CT due to hypersensitivity to iodinated contrast media, renal failure estimated by a glomerular filtration rate < 30 ml/min or pregnancy.

According to literature data, the proportion of vascular invasion in a similar population has been estimated between 25 and 30%, which meant that 50 to 60 patients were required for the 15 MVI events needed to achieve univariate analysis of the radiological variables [27, 28].

Seventy-five patients were eligible but 14 patients were excluded because they received radiofrequency ablation instead of surgery (n = 2), locoregional (n = 11), or systemic (n = 1) HCC treatment. Five of the 61 remaining patients were excluded for the following reasons: biphenotypic tumor on histological analysis (n = 3) or failure to cover HCC lesion by the CT perfusion volume (n = 2) (Fig. 1). Thus, clinical data, alpha-fetoprotein (AFP) serum values and liver transient elastography results were obtained for the 56 included patients. All patients had a single solitary nodule, except for two patients with two nodules, which amounts to a total of 58 nodular-type HCC lesions. All patients were followed postoperatively to detect early tumor relapse within 3 years. For this, AFP serum levels were monitored and dynamic CT or MRI were performed every 6 months.

Fig. 1
figure 1

Flowchart of the study. *HCC hepatocellular carcinoma, DECT dual-energy CT

Full size image

CT protocols

Patients underwent perfusion CT, followed by multiphase abdominal DECT with a 64-section multi-detector CT unit (Revolution HD GSI, GE Medical Systems Healthcare) ten minutes after the end of the perfusion image acquisitions. Unenhanced liver CT was first performed with a 140 mm-wide scan volume overlaying the HCC and the portal vein. Five seconds after administration of 45 ml of a non-ionic contrast agent (Iomeprol, 350 mg iodine /ml, Iomeron 350, Bracco Imaging) at a flow rate of 4 ml/sec followed by 20 ml of saline solution, perfusion CT was repeatedly performed in shuttle mode during normal breathing, with 25 volume acquisitions every 1.7 s over the first 42 s and then three volume acquisitions of 1.7 s every 30 s which amounts to a total examination time of 145 s (scan parameters: Table 1). Once perfusion acquisitions were complete, an arterial DECT phase (using bolus tracking software, acquisition began 6 s after the attenuation in abdominal aorta reached the threshold of 180 HU), a portal phase including a thoracic acquisition (60 s), and a delayed phase CT (3 min), were performed with administration of the same contrast agent (2 ml/kg minus the 45 ml already used for perfusion CT) at a flow rate of 4 ml/s (scan parameters: Table 1).

Table 1 CT protocols
Full size table

Perfusion CT and dual-energy CT data analysis

Results were processed at a workstation (Advantage Windows server 7.0, GE Medical Systems) by a radiologist (M.L with 15 years of experience in abdominal imaging) who was blinded to clinical and pathological data, using commercial CT perfusion software (GE CT perfusion 4D) for quantitative analyses. An integrated motion correction algorithm was first applied to all perfusion CT images and a deconvolution model with a dual vascular input algorithm was used to generate functional maps: Hepatic arterial Blood Flow (HaBF, in ml/100 g/min), Hepatic arterial flow (HAF, ratio between 0 and 1), Blood flow (BF, ml/100 g/min), Blood volume (BV, ml/100 g), Mean slope of increase (MSI, in HU/s), Tmax (seconds), Time to peak (TTP, seconds), Mean Transit Time (MTT, seconds), and Permeability surface-area product (PS,%). Iodine concentration maps were then automatically generated from arterial phase DECT images (GSI viewer, GE healthcare). The largest cross-sectional area of HCCs with the highest visual liver-to-lesion contrast were selected on perfusion CT images and regions of interest (ROI) were manually placed around the HCCs. Finally, all ROIs were automatically copied on parametric perfusion maps, iodine maps (arterial iodine concentration, IC (100 µg/ml)), and on arterial phase images (arterial density value, AD, (HU)) (Fig. 2). Similarly, standardized ROIs were drawn on the non-tumorous liver parenchyma and large hepatic vessels were avoided.

Fig. 2
figure 2

Perfusion CT map (hepatic arterial blood flow) showing a manually placed region of interest (ROI) around the hepatocellular carcinoma (a). The ROI was automatically copied on to iodine maps (b) and arterial phase images (c)

Full size image

Patient exposure and radiation burden

Exposure parameters (DLP, mGy cm) were collected for perfusion, DECT, unenhanced, portal and delayed CT scan sequences using the Radiation Dose Monitor (RDM, Medsquare) Dose Archiving and Communication System (DACS).

Histopathologic analysis

Pathological examination of surgical specimens was simultaneously reviewed by two experienced liver pathologists (A.LB and C.G with 5 and 20 years’ experience, respectively) and when interpretation was difficult a consensus was reached. Each nodule was analyzed for the following: (a) size, (b) presence of a capsule, (c) capsular invasion, (d) tumor differentiation according to a three-grade system (World Health Organization, WHO, 2019), (e) MVI defined as a tumor within a vascular space circled by endothelium, identified as intratumoral, peritumoral, or both, (f) satellite nodules, (g) the presence of an MTM component > 40%, (h) percent of tumor necrosis, (i) immunohistochemical expression of EpCAM (epithelial cell adhesion molecule) and AFP. In addition, the non-tumoral liver was assessed for steatosis (%) and fibrosis according to the METAVIR score.

Statistical analysis

Univariate logistic regression analyses were performed on tumor size and on each quantitative radiological perfusion parameter for testing MVI as outcome of the model. Concerning the radiological parameters, all data were integrated into two unsupervised multivariate models: principal component analysis (PCA) and unsupervised heatmap classification. First, PCA was performed with the FactoMineR R-package [29], only based on quantitative radiological parameters. Histological parameters were used as outcomes of the model. Significant Pearson correlations to pathological outcome parameters were extracted from PCA axes. In a second step, following the logarithm transformation of the radiological parameters, unsupervised classification was performed on a quantitative imaging heatmap using the ward_D2 method and Euclidean classification distances with illustrative integration of the histological parameters which were found to be significant using a PCA approach. The resulting unsupervised multivariate classification model was revealed as a heatmap using the pheatmap R-package [30]. When comparing mean values between different clusters for predictive radiological parameters, optimum binary cutoffs were selected for each significant parameter. Univariate survival analysis, censored at time of relapse, was performed using the log-rank test method for each stratification. Multivariable cox survival analysis, still censored at time of relapse, was performed with parameters that were identified as significant from univariate analysis (survival R-package). Statistical analysis was performed using a R-software (version 3.6.1) [31]. p ≤ 0.05 was considered to be significant.

Results

Clinical and pathological characteristics of patients

Patient characteristics are summarized in Table 2. Most of the 56 patients were men (44/56, 78%), median age 67 years old (range 31–84) with chronic viral infection (21/56, 37.5%). All patients were classified according to the Barcelona-clinic liver cancer staging system as stage A. Two of the 56 patients had two tumors, which amounts to a total of 58 nodular-type HCC lesions [median size of 42.3 mm (range 20–140 mm)].

Table 2 Patient characteristics
Full size table

The pathological features of the tumors are summarized in Table 3. Most tumors were moderately differentiated (36/58, 62%). Intratumoral and peritumoral MVI were present in 20/58 (34%) and 25/58 (43%) of the tumors, respectively. The MTM component > 40% was observed in 11/58 (19%) of the tumors. A capsule was found in 46/58 (79%) and capsular invasion was present in 21/46 (45%) of these cases. The non-tumoral liver parenchyma was scored F3/F4 in 38/56 (68%) of the patients. The size of HCCs was significantly correlated to MVI (p = 0.008).

Table 3 Pathological features of hepatocellular carcinoma (HCC) nodules
Full size table

Tumor relapse occurred in 19/56 (34%) patients within a period of 6 month up to 3-year.

Perfusion CT and dual-energy CT quantitative analysis

Tumor and tumor-free liver parenchyma could be discriminated based on analysis of the eleven quantitative CT perfusion and DECT parameters (Table 4, Fig. 3). The AD, IC, HaBF, HAF, and MSI were significantly higher in HCC lesions than in the tumor-free liver parenchyma while BV, Tmax, TTP, and MTT were significantly lower (p < 0.001). Univariate logistic regression performed on quantitative perfusion radiological parameters with MVI (intra and peritumoral MVI) did not reveal any significant association between the two. However, AD and the IC were significantly correlated to MVI (108 HU in HCC lesions with MVI versus 132 HU in HCC lesions without MVI, p = 0.001 and 27 × 100 µg/ml in HCC lesions with MVI versus 33 × 100 µg/ml in HCC lesions without MVI, p = 0.015, respectively) (Table 5).

Table 4 Dual-energy CT (DECT) and Perfusion CT-derived quantitative parameters
Full size table
Fig. 3
figure 3

Dual-energy CT (DECT) and Perfusion CT images in a 31-year-old man with hepatocellular carcinoma in hepatic segment VI. Axial images of a arterial phase, b iodine concentration, derived from multiphasic contrast-enhanced DECT. Axial parametric maps of c hepatic arterial blood flow (HaBF), d hepatic arterial flow (HAF), e mean slope of increase (MSI), and f permeability surface-area product (PS) obtained from perfusion CT. Liver resection specimen (hemato-eosin-saffron ×80). g the tumor shows a macrotrabecular liver architecture and microvascular invasion in peritumoral portal tract (star)

Full size image
Table 5 Imaging parameters of HCC without microvascular invasion (MVI−) and HCC with microvascular invasion (MVI+)
Full size table

The radiation burden from perfusion CT and GSI spectral imaging showed that the mean ± standard deviation DLP values for hepatic perfusion CT, arterial DECT, unenhanced, portal, and delayed CT scan sequences were 1309 ± 143, 462 ± 166, 286 ± 147, 666 ± 283, and 316 ± 160 mGy cm, respectively.

Prediction of histological parameters by the radiological model

The combination of the 11 quantitative parameters using PCA showed that a multivariate radiological model were predictive of the following histological features of HCC: a major WHO grade (p = 0.01), intratumoral MVI (p = 0.004), peritumoral MVI (p = 0.04), a MTM component > 40% (p = 0.02), and capsular invasion (p = 0.02) (Fig. 4) with respective loadings on the first three dimensions: PCA1 (35.98%), PCA2 (23.10%), PCA3 (17.36%).

Fig. 4
figure 4

Principal component analysis (PCA) projection of the multivariate radiological model with significant histological parameters. *WHO  World Health Organization

Full size image

However, the radiological parameters were not found to be correlated to tumor necrosis, the presence of satellite nodules, immunohistochemical expression of EpCAM and AFP, or AFP serum values.

In addition, PCA showed that a non-tumoral multivariate radiological model could predict steatosis and fibrosis in the tumor-free liver parenchyma (steatosis > 20% and fibrosis F0-1-2 versus F3-F4 stratification scores, p = 0.03 and p < 0.001, respectively).

HCC classification using radiological model with selected pathological annotations

A heatmap classification of HCC nodules was obtained by the multivariate radiological model with selected pathological annotations (Fig. 5). Unsupervised classification showed that the association of these 11 quantitative parameters reflected tumor heterogeneity and that patients could be stratified into three main clusters (R1, R2, and R3). R1, the cluster on the left, showed a low recurrence risk due to the absence or a very rare occurrence of intra- and/or peritumoral MVI or MTM-subtype. The clusters R2 on the right and R3 in the middle which included most of the tumors, showed a higher risk of recurrence since most of them proved positive for intra- and/or peritumoral MVI and a MTM component > 40%. Interestingly, the left-hand R1 cluster had a low level of PS. In contrast, the right-hand R2 cluster and the R3 middle cluster had medium and high levels of PS, respectively.

Fig. 5
figure 5

Unsupervised classification on heatmap of 11 quantitative imaging parameters with pathological annotations: heatmap of cells representing radiological parameter quantifications with adjacent clustering performed by ward_D2 method and Euclidean distances. This clustering allowed patient stratification in three main clusters: R1on the left (characterized by a low level of PS), R2 on the right and R3 in the middle. *HAF hepatic arterial flow, MSI mean slope of increase, HABF hepatic arterial blood flow, BF blood flow, AD arterial density, PS permeability surface-area product, MTT mean transit time, BV blood volume, TTP time to peak, peri_MVI peritumoral microvascular invasion, intra_MVI intratumoral microvascular invasion, Cap_Inv capsular invasion, WHO_grade major World Health Organization grade, MTM_subtype macrotrabecular-massive subtype

Full size image

Association of early recurrence with quantitative radiologic parameters

Quantitative radiological parameters were ranked individually compared to relapse status (19/56, 34%) to evaluate their respective optimal threshold of stratification for survival analysis (Table 6). Four of the 11 quantitative radiological parameters were found to be significant on univariate analysis by the log-rank test. The thresholds for the parameters were 120 for AD (log-rank p value = 0.009), 22.79 for PS (log-rank p value < 0.001), 7.7 for Tmax (log-rank p value = 0.02), and 23.2 for TTP (log-rank p value = 0.01). To test the independence of these prognostic parameters, a Cox multivariate model was built including AD, PS, Tmax, and TTP. In this model, only PS was found to be an independent and significant parameter to predict the risk of relapse (Cox p value = 0.004). According to the Log-rank analysis, if the PS value is > 22.79, the median estimated time to relapse is 1 year, otherwise the median time to relapse is > 2 years (Fig. 6).

Table 6 Univariate and Multivariate analyses of radiological relapse risk factors
Full size table
Fig. 6
figure 6

Log-rank analysis on independent PS parameter censored at time of relapse

Full size image

Discussion

This study shows that an imaging profile using a combination of quantitative tumoral CT parameters obtained from perfusion CT and DECT can predict the presence of peritumoral and intratumoral MVI, capsular invasion, tumor grade, and the MTM-subtype in HCC. Moreover, this radiological profile can be used to determine tumor heterogeneity and identify different sub-groups of HCC associated with specific histological features. PS appears to be the only independent prognostic parameter of the quantitative tumoral CT parameters that were tested.

By comparing our experimental quantitative CT perfusion and DECT results with the results in the literature, we obtained recognized values allowing us to build our radiologic model [21, 26, 32]. The imaging profile obtained with the 11 quantitative tumoral CT parameters was predictive of prognostic histological factors in HCC nodules. In fact, PCA identified a homogeneous, well-characterized group with significant invasive histological parameters (MVI, capsular invasion, MTM-subtype) that were also associated with WHO grade progression. Thus far, the only histological parameter identified using CT perfusion was reported by Sahani et al., who showed a significant difference in CT perfusion parameters between well-differentiated HCC and other grades (p < 0.05) [21]. Other approaches based on image descriptors have also been proposed for the prediction of MVI, such as radiogenomic venous invasion signatures on preoperative contrast-enhanced CT by Banerjee et al., or radiomic models based on texture analysis by Xu et al. and Ma et al., and a more recent model based on volumetric iodine quantification using DECT by Kim et al. [33,34,35,36]. But, to the best of our knowledge, there are no data on CT perfusion and MVI in HCC.

In our study intertumoral heterogeneity could be stratified with the quantitative imaging profile. Stratification helped identify patients at a high or low risk of relapse after hepatic resection, which could be used to determine specific therapeutic and management strategies in clinical practice. The performance of statistical models based on radiomic analysis has been shown to be good for the prediction of recurrence in early stage of HCC by Ji et al. [37]. The correlation between radiogenomics and the phenotype of aggressive disease and certain gene expression signatures on qualitative imaging has also been investigated by Taouli et al. [38]. However, none of these models are used in routine clinical practice. Conversely, the present study found that the preoperative tumor value of PS was an independent predictor of the risk of relapse. Interestingly, PS is a well-known parametric map reflecting tumor angiogenesis at varying degrees of maturation, which has mainly been used as a tumor response biomarker after targeted therapy, but not for preoperative prognosis [21, 39, 40]. It could be hypothesized that high PS values (> 22.79) suggest aggressive behavior in HCC due to the MVI or MTM-subtype.

The quantitative imaging profile could also be used to assess the non-tumoral liver and predict liver steatosis (threshold > 20%) and liver fibrosis (F0/F1/F2 fibrosis stages versus F3-4 stages). These same results were reported by Van Beers et al. in patients with cirrhosis [41]. In addition, Ronot et al., showed that perfusion CT can discriminate patients with early stages of liver fibrosis [42]. Incorporating the noninvasive assessment of liver fibrosis and steatosis could help extend the selection criteria for HCC resection.

The patients in our study were exposed to higher radiations doses than with a standard dynamic study, which may be of concern. However, when the DLP values registered in our study are compared to national Diagnostic Reference Levels for chest-abdomen-pelvis CT (DLPref = 750 mGy cm) only perfusion CT acquisitions were found to exceed the limit, while arterial DECT were within reference levels. Nevertheless, we feel that these perfusion CT radiation doses can be justified and compensated by the benefits and added value of the diagnostic information. Indeed, PS preoperative data, provided by perfusion CT for identifying aggressive HCC may be beneficial for both therapeutic and prognostic purposes. This protocol must now be optimized in further studies by reducing the radiation dose while keeping a good quality image.

Our study definitively has some limitations. First, the number of patients was low and only surgically confirmed HCCs were included, which could create a selection bias. Secondly, a 64-section multi-detector was used, while a 256-detector row scanner might have improved the performance of perfusion analysis. Moreover, DLP values for arterial acquisition could be improved by adjusting tube current. Thirdly, there could be a potential element of confusion in DECT parameters (AD and IC), resulting from the residual contrast from previous injections of perfusion CT. Advanced fibrosis could also affect quantitative parameters of liver perfusion. Fourthly, 2D ROIs were used instead of volumes-of-interest, because we believe that the selected surface reflects the heterogeneity of HCC. Fifthly, even though a strong inter-reader agreement has already been reported on the basis of a standardized CT protocol with the same CT perfusion software, we have to admit that the inter-reader agreement was not evaluated in the measured quantitative parameters [24]. Finally, the reproducibility of this quantitative imaging profile was not evaluated in an external validation cohort, which means that further multicenter studies with other CT manufacturers are required, in order to validate this imaging profile before it can be used in routine clinical practice [43, 44].

In conclusion, this study presents an imaging profile using a combination of quantitative perfusion CT and DECT parameters to diagnose histological MVI and MTM-subtype in patients with HCC prior to surgery. The preoperative tumor value of PS is an independent predictor of patients at a high risk of relapse prior to treatment.