Introduction

Hepatocellular carcinoma (HCC) represents the most frequent type of primary liver cancer and ranks as the third leading cause of cancer-related death worldwide [1]. Surgical resection is the backbone of curative treatment for HCC. Although the surgical operation of HCC has been extremely developed, the prognosis of HCC is still very poor [2]. The recurrence after operation remains a major obstacle to improve survival and the 5-year recurrence rates are up to 70%, even in patients with single small HCC (≤ 2 cm) [3]. One of the main reasons is microvascular invasion (MVI). MVI, a histopathological feature of micro-metastases, can be observed in 15.0%-57.1% of surgical specimens and stands as a key predictor for early recurrence and poor prognosis [4, 5]. In addition, the presence of MVI has been an essential factor influencing the long-term prognosis of HCC patients underwent liver transplantation [6]. However, the impact of MVI has largely been disregarded by clinicians as it can only be found by histologic examination of postoperative specimens. As MVI represents an aggressive behavior of HCC [5], individualized surgical management for patients should be formulated according to the risk–benefit assessment of MVI before surgery. For patients with high risk of MVI, anatomical subsegmentectomy or partial hepatectomy with wide margin is recommended under better liver function, which could decrease the risk of recurrence [7]. Therefore, establishing a highly accurate risk model for predicting MVI can help to stratify high-risk individuals, optimize clinical decision making and improve prognosis.

Over the last decades, numerous efforts have been taken to estimate MVI prior to surgery. Cucchetti et al. utilized four clinical common variables, including alpha-fetoprotein (AFP), tumor number, size and volume, to develop an artificial neural network (ANN) model that accurately identified 91% of MVI cases in testing group [8]. However, this ANN model requires specialized computer software, which potentially limits its utility in routine clinical application. In contrast, the nomogram might be a convenient option and provide highly accurate and individualized risk estimate [9]. For hepatitis B virus–related HCC patients, Li et al. constructed a nomogram by incorporating contrast-enhanced magnetic resonance imaging (MRI) data with hematology data, which achieved relatively good predictive accuracy for MVI [

Results

Baseline characteristics of patients

During the study period, 158 HCC patients form institution A and 53 HCC patients form institution B met the inclusion criteria were enrolled. The former was the training cohort with the median follow-up time of 23.9 months, and the latter was the validation cohort with the median follow-up time of 27.1 months. Pathological confirmed MVI was found in 101 (47.9%) individuals of all patients. The detailed clinical, adipose as well as muscle related variables were shown in Table 1, respectively.

Table 1 Baseline characteristics of HCC patients
Full size table

Data dimensionality reduction and screening

LASSO regression analysis was performed to identify the most crucial variables linked with the risk of MVI. To guarantee precision, the algorithm's iterations were set to 1000. In addition, cv.glmnet function was used to perform tenfold cross validation to diminish the possibility of overfitting. The distribution of LASSO regression coefficients and cross-validation plots were shown in Fig. 2. Dotted vertical lines indicated the value of lambda.min and lambda.1se. Finally, 18 variables (age, tumor size, CTP, cirrhosis, HBsAg, AFP, TB, resection method, SAT density, VAT area, VAT density, VATI, VSR, IMAT area, IMATI, SM area, SM density, SMI) with nonzero coefficients were selected according to the value of lambda.min (λ = 0.03139931).

Fig. 2
figure 2

Selecting influence factors by using LASSO regression model. a LASSO coefficient curves for 28 predictors. b Identification of the best punishment coefficient lambda in LASSO regression model

Full size image

Logistic regression analysis, construction, and validation of nomogram

The above 18 variables were subjected to univariate logistic analysis, and then 8 variables (age, tumor size, cirrhosis, VAT density, IMAT area, IMATI, SM area, SMI) with P-values of less than 0.1 were incorporated into multivariate analysis. Except for IMAT area and SMI, the remaining 6 variables [age (OR = 0.214; 95%CI = 0.088–0.522; P = 0.001), tumor size (OR = 2.359; 95%CI = 1.115–4.989; P = 0.025), cirrhosis (OR = 2.518; 95%CI = 1.077–5.887; P = 0.033), VAT density (OR = 3.468; 95%CI = 1.533–7.847; P = 0.003), IMATI (OR = 4.783; 95%CI = 2.058–11.112; P < 0.001), SM area (OR = 0.295; 95%CI = 0.120–0.725; P = 0.008)] were all significant independent predictors for MVI. The detailed data was shown in Table 2.

Table 2 Univariate and multivariate logistic analysis of MVI presence for HCC patients
Full size table

These Six independent predictors were utilized to form an MVI-predicting nomogram (Fig. 3a). The nomogram provided good accuracy in estimating the risk of MVI, with C index of 0.79 (95%CI: 0.72–0.86), R2 of 0.35, discrimination index (D index) of 0.30, Brier score of 0.18, Emax of 0.05, Evag of 0.02 (Fig. 3b). Following 200 rounds of tenfold internal cross-validation, there was little variation in the values of the above variables (Table 3), with the correct C index of 0.79, R2 of 0.32, D index of 0.24, Brier score of 0.18, Emax of 0.30, Evag of 0.13. In addition, calibration curve exhibited good consistency between predicted MVI probability of nomogram and the actual probability.

Fig. 3
figure 3

Nomogram model for predicting MVI of HCC patients (a) and the calibration curve of the nomogram in training (b) and validation cohorts (c)

Full size image
Table 3 The changes variables after 200 rounds of tenfold internal validation
Full size table

In validation cohort, the nomogram presented a C index of 0.75 (95%CI: 0.62–0.89) for the evaluation of risk associated with MVI. Moreover, the fitting calibration curve was acceptable (Fig. 3c).

Clinical utility of the nomogram

The independent predictors of MVI included three clinical variables (age, tumor size and cirrhosis) and three imaging variables (VAT density, IMATI and SM area). These clinical and imaging variables were used separately to construct new clinical and imaging model, which were then compared with this nomogram to verify its clinical utility. The ROC curve indicated that the discriminative ability of the nomogram in predicting MVI was better than that of the imaging and clinical models, with AUC values of 0.791, 0.711, 0.673 and 0.716, 0.654, 0.633 in training and validation cohorts, respectively (Fig. 4a, b). In addition, the DCA of nomogram also exhibited a wider range of threshold probabilities in comparison to the imaging and clinical model (Fig. 4c, d).

Fig. 4
figure 4

ROC curve (a-b) and DCA (c-d) of the nomogram, imaging and clinical models in training and validation cohorts

Full size image

Risk stratification of MVI based on the nomogram and survival analysis

The optimal cut-off value of the nomogram scores in training cohort was determined to be 273.8, and the sensitivity and specificity for distinguishing the presence from absence of MVI according to this value were 79.7% and 63.3%, respectively (Supplementary Fig. 2). Based on the cut-off value, patients in two cohorts were divided into high (> 273.8) and low (< = 273.8) risk groups. Kaplan–Meier curves showed that in both training and validation cohorts, patients with high-risk of MVI had higher 2-year recurrence rate and shorter OS than patients with low-risk (Fig. 5). For patients with high MVI risk, wide-margin resection or anatomical resection could significantly improve the 2-year RFS, while only wide-margin resection could significantly improve the OS (Figs. 6 and 7). In contrast, for low-risk patients, neither surgical approaches, nor resection methods or surgical margins had impact on recurrence and OS (Supplementary Figs. 3, 4).

Fig. 5
figure 5

Kaplan–Meier curves of OS and 2-year RFS for patients with different risks scores in training cohort (a, b) and validation cohort (c, d)

Full size image
Fig. 6
figure 6

Kaplan–Meier curves of OS for high-risk patients under different surgical approaches, resection methods and surgical margins in training cohort (a-c) and validation cohort (d-f)

Full size image
Fig. 7
figure 7

Kaplan–Meier curves of 2-RFS for high-risk patients under different surgical approaches, resection methods and surgical margins in training cohort (a-c) and validation cohort (d-f)

Full size image

In order to facilitate the clinical use of the nomogram, we have provided an online version accessible via the hyperlink: https://sdmxc.shinyapps.io/MVInom/. By setting variables greater than or equal to the cut-off value to 1 and those less than it to 0, the predicted probability of MVI for patients can be quickly obtained.

Discussion

In the present study, we showed that adipose and muscle tissues-related variables, including VAT density, IMATI and SM area, were capable to serve as independent predictors for MVI. Furthermore, we incorporated the above variables with three common clinical variables (age, tumor size, cirrhosis) to develop an MVI-predicting model which exhibited good consistency and high accuracy. In the process of validating the clinical utility, our nomogram was also superior to the models only based on imaging or clinical variables. Compared with other radiomics nomograms [11, 12], we only need one plain CT image at L3 level for data extraction, without the need to conduct complex three-dimensional tumor target delineation at various phases of enhanced scanning, which will bring heavy computational burden. Recently, some researchers have employed liver biopsy specimens to analyze gene expression or even radio-genomics for predicting MVI [31, 32]. Despite its good predictive performance, tissue biopsy for the diagnosis of HCC is unnecessary for patients with typical imaging findings according to some clinical guidelines [33, 34]. Additionally, invasive procedure may lead to potential risks such as tumor seeding and fatal bleeding, which outweigh the benefits [34, 35].

In the previously published MVI-predicting nomograms, tumor size, cirrhosis and age has been confirmed as risk factors for vascular invasion in HCC. Pawlik et al. reported that the incidence of MVI was almost twice as high in tumors larger than 5 cm (61%) when compared to smaller tumors (32%), and it increased continuously with increasing tumor size [36]. Cirrhosis patients were frequently present with alterations in hemodynamic and blood microenvironment. On one hand, secondary hypersplenism could decrease platelet counts and reduce portal blood flow to the liver, which was conducive to the formation of MVI [10]. On the other hand, the increase of von Willebrand factor and other procoagulants, as well as the decrease in generation of antithrombin in patients with cirrhosis, could cause an imbalance in the coagulation and fibrinolytic systems [37, 38]. The mechanisms mentioned above could both serve as stimuli for thrombosis and vascular invasion. In current study, younger HCC patients had a higher incidence of MVI, which was in consistent with another research [12]. This observation might be attributed to the high aggressiveness of the tumor in young patients; however, the exact underlying mechanism remains to be investigated by future studies [39].

For body composition-related variables, metabolic theory could support our findings. Because most fatty acids in VAT are transported via the portal vein, higher degree of visceral obesity will expose the liver to high concentrations of fatty acids and glycerol. This, in turn, leads to a series of liver dysfunctions, such as reduced hepatic insulin extraction, increased hepatic glucose production, and ultimately results in type 2 diabetes and insulin resistance [40]. In addition, accumulation of visceral adipose is also linked with the upregulation of aquaporin adipose, an adipocyte-specific glycerol channel, leading to the increased glycerol secretion from VAT [41]. The liver-specific aquaporin protein AQP-9 could absorb excess glycerol and convert it into glycerol phosphate through glycerokinase, which is also one of the mechanisms leading to hyperglycemia [42]. Hyperglycemia alters the size and composition of the basement membranes in micro and large vessels, which in turn increase vascular permeability and fragility, and make it easier for tumors to invasion vascular and metastasize [43,44,45]. Furthermore, with continues accumulation of lipids, adipose tissues will undergo complex process of remodeling, and the secretion of adipokines, such as leptin, adiponectin, interleukin, vascular endothelial growth factor (VEGF), and hypoxia-inducible factor (HIF) will significantly increase [46], which could induce angiogenesis by stimulating endothelial cell proliferation, migration, and tube formation [47].

SM area provides information on changes in muscle structure, while IMAT indicates alterations in the content of intra- and inter-myocellular adipose [48]. The former parameter denotes muscle quantity and is served as an assessment metric for sarcopenia, and the latter parameter reflects muscle quality and is served as an evaluation index for myosteatosis [20]. Recent studies have established that sarcopenia is positively associated with evaluated all-cause mortality in HCC patients [49, 50]. Though the detailed mechanism has yet to be fully elucidated, however, there are two generally accepted hypotheses. Depletion of muscle could lead to a reduction in the secretion of specific cytokines, such as insulin-like growth factor (IGF)-1, which is associated with advanced clinicopathological variables of HCC [51, 52]. In addition, sarcopenia could also result in the impairment of certain protein hydrolysis systems, such as the tumor necrosis factor (TNF)-α system, thereby facilitating the process of tumor migration and invasion [53].

To date there have no studies that illuminate molecular mechanisms between myosteatosis and tumor angiogenesis. However, pathological changes in muscle have a strong correlation with chronic liver disease [54]. Montano-Loza et al. studied a cohort of cirrhotic patients undergoing liver transplant evaluation, in which the prevalence of muscular steatosis was greater than 50% [55]. In cirrhosis patients, the absorption of lipids and fat-soluble vitamins may be impacted as a result of several physiological changes, such as nausea, anorexia, increased intra-abdominal pressure, and impaired gut motility. Malnutrition promotes the catabolism of muscle proteins, leading to depletion of muscle mass [56, 57]. In addition, chronic hepatic inflammation plays a significant role in promoting systemic inflammation, which in turn also accelerates the catabolism of muscle proteins [58]. There have been studies shown that liver disease not only triggers muscle atrophy and changes its structure, but the muscle in turn promotes the further development of liver disease [54, 59]. Therefore, there may exist an interaction between liver disease and muscle tissue, which exerts an influence on the IMAT, and consequently modulates the expression of downstream adipokines and myokines. Such a dynamic interaction and these cytokines may contribute to the tumor invasion and progression of liver disease.

There are several limitations in our study. Firstly, the limited number of patients impose inherent restrictions on our findings, thus larger external datasets are needed to validate and refine it. Secondly, a prospective study is required to further confirm the reliability of the nomogram. Thirdly, MVI grade was not taken into account in the MVI positive group. Fourthly, this nomogram model is constructed to predict the risk of MVI. Since macrovascular invasion is also a common phenomenon in HCC and is known to be an important predictor of poor survival outcomes [60, 61], whether this nomogram model is fitted for the predication of macrovascular invasion in HCC deserves further study.

In addition, some baseline characteristics showed difference between the two cohorts, and it’s inevitable as the process of data collection in these two hospitals were conducted independently. However, these differences had minimal impact on our results. Firstly, data selection and model construction were both performed within training cohort. Two hundred rounds of internal tenfold cross-validation method was used to verify the stable of our model, which could reduce overfitting of result and maximize the utilization of the data. Secondly, although split sample approach could create two cohorts with no differences [62], it could lead the model being unstable, that is, the performance evaluation was correlated with the partition of the training and validation cohorts. Different partitions might yield different estimates of the accuracy [14, 63]. Therefore, we chose original cohort distribution to ensure the randomness and independence of the data and increase generalizability of the model. In fact, we also verified our model by propensity score matching method to match the baseline characteristics data of the validation cohort and the training cohort, and it still demonstrated good predictive capability (data not shown). This suggests that our model has good generalizability and can be extrapolated to patients in other hospitals.

Conclusion

By combining 6 preoperative independently predictive factors of MVI, a nomogram is constructed. This model provides an optimal preoperative estimation of MVI risk in HCC patients. For patients with higher MVI scores, anatomical hepatectomy with wide margins may be recommend to reduce early recurrence and extend survival.