Correlation study of 18F-FDG PET/CT metabolic parameters, heterogeneity index, and microvascular invasion, and its nomogram potential in predicting microvascular invasion in liver cancer before liver transplantation

Abstract

Background

Hepatocellular carcinoma (HCC) is a highly prevalent malignant tumor worldwide, with Chinese patients accounting for more than 50%. Microvascular invasion (MVI) is a significant risk factor for postoperative recurrence of HCC. ¹⁸F-fluorodeoxyglucose PET/computed tomography (¹⁸F-FDG PET/CT), as a hybrid imaging modality integrating metabolic information from PET with anatomical details from CT. This combined approach enables simultaneous assessment of glucose metabolism and structural features. It can also evaluate tumor biological behavior through metabolic parameters and heterogeneity characteristics.

Objective

To explore the predictive value of ¹⁸F-FDG PET/CT metabolic parameters and heterogeneity index for MVI in HCC patients before liver transplantation and to construct a nomogram prediction model.

Methods

A retrospective study involving 177 HCC patients who underwent liver transplantation (100 MVI-positive cases and 77 MVI-negative cases) was conducted to analyze the correlation between clinical characteristics, PET/CT metabolic parameters (SUVmax, SUVmean, TLG, and TLR), and heterogeneity parameters (COV and HI) with MVI. Independent predictors were identified using univariate and multivariate logistic regression, and a nomogram model was constructed. The model's performance was evaluated using calibration curves and ROC curves.

Results

Univariate analysis showed significant differences in PIVKA-II, SUVmax, TLG, TLR, COV, and HI between the two groups (all P < 0.05). Multivariate analysis indicated that PIVKA-II (OR = 1.000, P = 0.042), TLG (OR = 0.999, P = 0.024), HI (OR = 1.022, P < 0.001), and TLR (OR = 1.618, P = 0.031) were independent predictors of MVI. The area under the ROC curve (AUC) of the combined model reached 0.815 (95% confidence interval: 0.754–0.876), significantly better than any single parameter. The nomogram calibration curve showed a high consistency between predicted probabilities and actual observed probabilities (mean absolute error = 0.025).

Conclusion

The integration of PET/CT-derived parameters—specifically TLG (metabolic burden), HI (heterogeneity), and TLR (tumor-to-liver contrast)—with serum PIVKA-II provides a robust tool for preoperative MVI prediction in HCC patients undergoing liver transplantation. The validated nomogram model (AUC = 0.815) outperforms individual parameters, offering a reliable basis for clinical decision-making.

Hepatocellular carcinoma (HCC) ranks sixth in the incidence of malignant tumors worldwide and fourth in tumor-related mortality. China is a high-incidence country for HCC, with new cases and deaths each year accounting for more than 50% of the global total [1]. Chronic hepatitis B or C is the main cause of HCC, accounting for 80% of HCC cases globally. The main treatment options for HCC include interventional surgery, surgical resection, or liver transplantation [2]. Liver transplantation is the best choice for radical treatment of early stage HCC [3], but about 20–30% of patients experience recurrence [4]. Therefore, stricter selection criteria are needed to choose liver transplantation candidates to ensure good outcomes. Reports indicate that the most important clinical features predicting early recurrence after liver transplantation include tumor diameter, number of lesions, histological grading, satellite nodules, lymph node metastasis, and microvascular invasion (MVI) [5,6]. Moreover, MVI is an important predictor of recurrence after surgical resection or liver transplantation for HCC. Most early recurrences of HCC are pathologically confirmed as MVI-positive postsurgery [7]. Clinically, there is a need for reliable prediction of the presence of MVI to reduce the likelihood of postoperative HCC recurrence, better select patients who could benefit from liver transplantation, and potentially improve their overall survival prognosis. MVI is a pathological feature described in HCC resection specimens under a microscope, characterized by clusters of cancer cells appearing within endothelial cell-lined vascular lumens. It primarily involves the small branches of the portal vein, small venous branches in fibrous septa or tumor capsules, and the central veins of the nontumorous liver tissue [8]. MVI is mainly detected through immunohistochemistry and pathological analysis of postoperative tissue samples, which are invasive and limited [9].

In recent years, many reports have assessed MVI preoperatively by extracting PET/CT image features. To date, ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) is the most widely used tracer in tumor applications. Some studies [10] have indicated that ¹⁸F-fluorodeoxyglucose PET/computed tomography (¹⁸F-FDG PET/CT) can be used for MVI prediction, but the results have application limitations, and its clinical value still needs further exploration. Tumor heterogeneity indices such as the coefficient of variation (COV) and heterogeneity index (HI) have been established as valuable tools for characterizing metabolic heterogeneity in malignancies including lung, breast, and pancreatic cancers [11–13]. However, their utility combined with metabolic parameters [total lesion glycolysis (TLG), tumor-to-liver ratio (TLR)] and serum biomarker (PIVKA-II) for preoperative prediction of MVI in HCC patients undergoing liver transplantation remains unexplored. This study aims to develop an integrated model leveraging these indices within a transplant candidacy assessment framework [12–16].

Materials and methods

General information

A retrospective inclusion of 177 HCC patients who underwent liver transplantation at our hospital from March 2019 to August 2024, with all cases pathologically confirmed postoperatively. Inclusion criteria: (a) Age >18 years; (b) Underwent ¹⁸F-FDG PET/CT within 2 weeks preoperatively; and (c) Complete clinical data and signed informed consent. Exclusion criteria: (a) Combined with other malignancies; (b) Incomplete case information. Finally, 162 males and 15 females were included, aged 18–75 years (median age 54 years). According to postoperative pathological MVI status, patients were divided into a positive group (n = 100) and a negative group (n = 77).

Instruments and methods

The ¹⁸F-FDG PET/CT examination was performed using a GE Discovery PET/CT (USA), model Discovery PET/CT 710. CT scan parameters were: tube current 150–250 mA, tube voltage 120 kV, pitch 0.8, slice thickness 3.75 mm. The imaging agent (¹⁸F-fluorodeoxyglucose injection, Zhejiang Andico Positron Emission Technology Co., Ltd., Jinhua, Zhejiang, China) is typically injected into the contralateral elbow vein after preparation, with radiochemical purity ≥95%. The dose is calculated based on body weight, with an adult injection dose of 3.7 MBq ¹⁸F-FDG/Kg. Image acquisition is typically performed within 60 min after the injection of the imaging agent, with the scanning range from the top of the skull to the upper part of the femur, scanning 6–7 beds, and the acquisition time for each bed is usually 2–3 min. Image reconstruction is usually done using the ordered subsets expectation maximization method. Attenuation correction is performed using CT data, and PET images are reconstructed using the ordered subsets expectation maximization iterative algorithm (two iterations, 24 subsets).

Image analysis

¹⁸ F-fluorodeoxyglucose PET/computed tomography image analysis

Diagnosed by two experienced physicians with senior titles from the nuclear medicine department. 3D reconstruction was performed on the AW4.6 workstation using computer-aided reporting software (GE Medical Systems S.C.S.283, Buc, France). Volumes of interest were drawn to obtain the maximum standardized uptake value (SUVmax), mean standardized uptake value (SUVmean), peak standardized uptake value (SUVpeak), SD of uptake values (SUVsd), liver SUVmax, and TLR. Volume parameters included metabolic tumor volume (MTV) and TLG. Calculate two tumor heterogeneity parameters: (a) COV, which is the COV, defined as SUVsd/SUVmean, as described in the literature [12]; (b) HI, which is the HI, defined as the absolute value of the slope of the linear regression of MTV at 40% SUVmax, 60% SUVmax, and 80% SUVmax thresholds, derived by modifying the methods proposed in previous literature [13,14].

Statistical methods

All statistical analyses were performed using SPSS 27.0 software. Categorical data were expressed as cases (%), and differences between groups were analyzed using the chi-square test. To determine whether the measurement data conformed to a normal distribution, the nonparametric Kolmogorov–Smirnov test was performed. Data that follows a normal distribution is expressed as ‘mean ± SD’, and intergroup comparisons are conducted using t-tests. Data that does not follow a normal distribution is expressed as ‘median (lower quartile, upper quartile)’, and the nonparametric Mann–Whitney U test is used. Multiple index joint diagnosis is achieved using logistic regression analysis, and a nomogram is constructed based on the regression model for individualized prediction of the probability of MVI occurrence. The consistency between predicted probabilities and actual observed probabilities is assessed using a calibration curve. The efficiency of diagnosing MVI positivity is analyzed using the receiver operating characteristic (ROC) curve. All tests for statistical significance are two-tailed, with P < 0.05 considered statistically significant. This study was conducted and reported in accordance with the TRIPOD (transparent reporting of a multivariable prediction model for individual prognosis or diagnosis) statement. A completed TRIPOD checklist detailing adherence to each reporting item is provided as Supplementary Material (Supplementary Table 1, Supplemental digital content 1, https://links.lww.com/NMC/A337).

Validation cohort design

To mitigate overfitting risks, the cohort was prospectively partitioned into a training set (n = 123, 70%) for variable selection and nomogram construction, and a validation set (n = 54, 30%) for independent testing of model performance. The partitioning was randomized, with no significant differences in baseline characteristics between the two sets (all P > 0.05), ensuring their comparability. Baseline characteristics and PET/CT parameters of the training cohort (n = 123) stratified by MVI status are provided (Supplementary Tables 2 and 3, Supplemental digital content 1, https://links.lww.com/NMC/A337).

Results

Clinical characteristics and laboratory data of patients

A total of 177 patients were included in this study, aged 18–75 years, with a median age of 54 years. Among them, 100 patients (56.5%) were detected with MVI. The level of PIVKA-II was significantly elevated in the MVI-positive group, with a statistically significant difference (P < 0.001); Differences in age, gender, alpha-fetoprotein (AFP), tumor size, number, differentiation, and staging did not reach statistical significance (P > 0.05) (Table 1).

Table 1.

Comparison of clinical characteristics of hepatocellular carcinoma patients before transplantation between microvascular invasion-positive and microvascular invasion-negative groups

Variable	MVI negative group (n = 77)	MVI positive group (n = 100)	P value
Gender			0.178
Male	68 (88.3%)	94 (94.0%)
Female	9 (11.7%)	6 (6.0%)
Age (years)	54.60 ± 9.226	52.18 ± 10.837	0.119
PIVKA-II (mAU/ml)	161.00 (37.00, 601.50)	1250.50 (203.25, 13 448.50)	<0.001
AFP (ng/ml)	29.80 (6.75, 166.00)	87.10 (7.53, 1003.10)	0.067
Diameter (mm)			0.297
<50	55 (71.4%)	64 (64.0%)
≥50	22 (28.6%)	40 (40.0%)
Number			0.055
Single	38 (49.4%)	35 (35.0%)
Multiple	39 (50.6%)	65 (65.0%)
Differentiation			0.556
High	8 (10.4%)	9 (9.0%)
Medium	59 (76.6%)	72 (72.0%)
Low	10 (13.0%)	19 (19.0%)
T stage			0.094
T1–T2	59 (76.6%)	65 (45.0%)
T3–T4	18 (23.4%)	35 (55.0%)

AFP, alpha-fetoprotein; MVI, microvascular invasion; PIVKA-II, abnormal prothrombin.

¹⁸F-fluorodeoxyglucose PET/computed tomography metabolic parameters

There are significant differences in SUVmax, SUVmean, SUVpeak, TLG, TLR, COV, and HI between the two groups (P < 0.005). HCC patients with high SUVmax (5.34 vs. 4.05; P < 0.001), high SUVmean (3.08 vs. 2.52; P = 0.002), high SUVpeak (4.45 vs. 3.58; P < 0.001), high TLR (1.83 vs. 1.24; P < 0.001), high TLG (333.50 vs. 95.00; P < 0.001), high COV (0.21 vs. 0.16; P = 0.006), and high HI (65.05 vs. 14.95; P < 0.001) have a higher proportion of MVI; there is no significant difference in MTV between the two groups (P = 0.192) (Table 2).

Table 2.

Comparison of ¹⁸F-fluorodeoxyglucose PET/computed tomography quantitative parameters

Variable	MVI negative group (n = 77)	MVI positive group (n = 100)	P value
SUVmax	4.05 (3.32, 5.52)	5.34 (4.13, 7.52)	<0.001
SUVmean	2.52 (2.17, 3.26)	3.08 (2.50, 4.08)	0.002
SUVpeak	3.58 (2.78, 4.53)	4.45 (3.48, 6.30)	<0.001
TLG (g/ml×cm³)	95.00 (29.00, 221.00)	333.50 (84.00, 923.25)	<0.001
MTV (cm³)	58.30 (21.25, 168.50)	88.45 (26.00, 208.77)	0.192
TLR	1.24 (1.09, 1.57)	1.83 (1.50, 2.55)	<0.001
COV	0.16 (0.13, 0.24)	0.21 (0.15, 0.26)	0.006
HI	14.95 (5.73, 30.70)	65.05 (17.40, 159.65)	<0.001

COV, coefficient of variation; HI, heterogeneity index; MTV, metabolic tumor volume; SUVmax, maximum standardized uptake value of the tumor; SUVmean, average standardized uptake value of the tumor; SUVpeak, peak standardized uptake value of the tumor; TLG, total lesion glycolysis; TLR, ratio of tumor SUVmax to liver SUVmax.

Univariate and multivariate logistic regression analyses for predicting microvascular invasion

Correlation analysis found that PIVKA II (P = 0.004), TLG (P < 0.001), MTV (P < 0.001), HI (P < 0.001), and TLR (P = 0.001) were significantly correlated with MVI, while SUVmax, SUVmean, SUVpeak, and COV did not reach statistical significance (P > 0.05). Incorporating the aforementioned significant differences, a multivariable logistic regression analysis including PIVKA Ⅱ, TLG, MTV, HI, and TLR showed that PIVKA Ⅱ [odds ratio (OR) = 1.000, 95% confidence interval (CI): 1.000–1.000, P = 0.042], TLG (OR = 0.999, 95% CI: 0.998–1.000, P = 0.024), HI (OR = 1.022, 95% CI: 1.010–1.034, P < 0.001), and TLR (OR = 1.618, 95% CI: 1.046–2.504, P = 0.031) are independent predictors of MVI (Table 3).

Table 3.

Univariate and multivariate logistic regression analysis of ¹⁸F-fluorodeoxyglucose PET/computed tomography quantitative parameters in the microvascular invasion-negative and microvascular invasion-positive groups

Variable	Univariate OR (95% CI)	P value	Multivariate OR (95% CI)	P value	VIF
Pivka-II	1.000 (1.000, 1.000)	0.004	1.000 (1.000, 1.000)	0.042	1.181
SUVmax	1.094 (0.993, 1.206)	0.069
SUVmean	1.167 (0.971, 1.402)	0.099
SUVpeak	1105 (0.984, 1.242)	0.092
TLG (g/ml×cm³)	1.002 (1.000, 1.003)	<0.001	0.999 (0.998, 1.000)	0.024	2.081
TLR	1.985 (1.315, 2.995)	0.001
COV	2.876 (0.175, 47.196)	0.459	1.618 (1.046, 2.504)	0.031	1.005
HI	1.019 (1.011, 1.027)	<0.001	1.022 (1.010, 1.034)	<0.001	2.224

This suggests that TLG and HI may partially capture metabolic information overlapping with SUV parameters. While SUVmax, SUVmean, and SUVpeak showed significant associations with MVI in univariate analysis (all P < 0.001), they were not retained as independent predictors in the multivariate model after adjusting for TLG, HI, TLR, and PIVKA-II. VIF (variance inflation factor) values indicate moderate collinearity for TLG (VIF = 2.081) and HI (VIF = 2.224), while PIVKA-II (VIF = 1.181) and COV (VIF = 1.005) showed low collinearity.

18F-FDG, 18F-fluorodeoxyglucose; CI, confidence interval; COV, coefficient of variation; HI, heterogeneity index; MVI, microvascular invasion; OR, odds ratio; SUVmax, maximum standardized uptake value of the tumor; SUVmean, mean standardized uptake value of the tumor; SUVpeak, peak standardized uptake value of the tumor; TLG, total lesion glycolysis; TLR, ratio of tumor SUVmax to liver SUVmax.

Prediction nomogram (illustration) of microvascular invasion probability

Based on multivariate logistic regression results, four independent risk factors predictive of MVI were incorporated into the nomogram presented in Fig. 1. Each variable corresponds to distinct point allocations quantifiable via the ‘Points’ scale at the top. For individual patients, specific values of PIVKA-II, TLG, HI, and TLR can be located on their respective axes to determine corresponding scores. Summation of these scores yields the ‘Total Points’, where higher values indicate elevated predictive values and increased MVI risk probability. The nomogram calibration curve (Fig. 2) demonstrates close proximity between the bias-corrected line and the Ideal line, indicating strong alignment between predicted probabilities and observed outcomes (calibration slope = 0.98). This signifies high calibration performance, with the mean absolute error (MAE = 0.025) in the bottom-right corner further supporting model stability and reliability. Collectively, the calibration curve validates the model’s robust predictive capability for MVI.

Fig. 1.

Nomogram for predicting MVI. MVI, microvascular invasion.

Fig. 2.

Calibration nomogram. Where the x-axis represents the predicted probability of MVI occurrence by the model, and the y-axis represents the actual observed proportion of MVI occurrence. The dashed line (ideal) is the ideal calibration line, the dotted line (apparent) is the calibration curve of the original data, and the solid line (bias-corrected) is the bias-corrected curve calculated using 1000 bootstraps. 18F-FDG, 18F-fluorodeoxyglucose; COV, coefficient of variation; HI, heterogeneity index; MTV, metabolic tumor volume; MVI, microvascular invasion; SUVmax, maximum standardized uptake value of the tumor; SUVmean, mean standardized uptake value of the tumor; SUVpeak, peak standardized uptake value of the tumor; TLG, total lesion glycolysis; TLR, ratio of tumor SUVmax to liver SUVmax.

Predicting microvascular invasion status

In the ROC curve analysis (Table 4, Figs. 1–3), the AUC values for different variables in predicting MVI all exceeded 0.7, indicating good discriminative ability. Among them, HI (AUC = 0.754, 95% CI: 0.684–0.825) and TLR (AUC = 0.744, 95% CI: 0.668–0.820) performed best in independently predicting MVI. PIVKA II (AUC = 0.710, 95% CI: 0.643–0.786) and TLG (AUC = 0.716, 95% CI: 0.641–0.791) also demonstrated good predictive efficacy. In terms of sensitivity and specificity, TLR has the highest sensitivity (0.760), meaning this indicator can more effectively identify MVI-positive cases, while HI has the highest specificity (0.831), indicating its stronger ability to distinguish MVI-negative cases. Regarding the best thresholds, the optimal cutoff for PIVKA is 511.00, for TLG is 244.750, for HI is 41.575, and for TLR is 1.492. The AUC of the combined variables reached 0.815 (95% CI: 0.754–0.876), which is significantly better than individual indicators, indicating that combining multiple variables can further improve the accuracy of MVI prediction. The sensitivity of the combined model is 0.740, and the specificity is 0.766, balancing the stability and reliability of the prediction. To mitigate overfitting risks, the dataset was randomly partitioned into a training cohort (70%, n = 123) and a validation cohort (30%, n = 54). Baseline characteristics were balanced between cohorts. Supplementary Table 2, Supplemental digital content 1, https://links.lww.com/NMC/A337 further details the comparison of clinical and PET/CT parameters by MVI status within the training cohort, confirming significant differences in PIVKA-II, TLG, HI, and TLR (all P < 0.05), consistent with the overall population.

Table 4.

ROC curve for predicting microvascular invasion in hepatocellular carcinoma

Variable	AUC (95% CI)	P value	Sensitivity	Specificity	Threshold
Pivka-II (mAU/ml)	0.710 (0.6435, 0.786)	<0.001	0.620	0.753	511.00
TLG (g/ml×cm³)	0.716 (0.641, 0.791)	<0.001	0.610	0.792	244.750
HI	0.754 (0.684, 0.825)	<0.001	0.620	0.831	41.575
TLR	0.744 (0.668, 0.820)	<0.001	0.760	0.711	1.492
Combined	0.815 (0.754, 0.876)	<0.001	0.740	0.766	/

AUC, area under the ROC curve; CI, confidence interval; HI, heterogeneity index; OR, odds ratio; TLG, total lesion glycolysis; TLR, tumor-to-liver ratio.

Fig. 3.

ROC curve for predicting microvascular invasion in hepatocellular carcinoma.

Data partitioning and validation

To mitigate overfitting risks and address reviewer concerns regarding model generalizability, we implemented rigorous cohort partitioning (Table 5).

Table 5.

Randomly divide the data into a 7:3 training set and a validation set

Variable	Tets data (n = 54)	Train data (n = 123)	P value
Gender			0.964
Male	50 (92.6%)	112 (91.1%)
Female	4 (7.4%)	11 (8.9%)
Age (years)	52.93 ± 9.381	53.37 ± 10.589	0.793
PIVKA-II (mAU/ml)	211.50 (44.50, 1314.00)	253.00 (70.00, 1743.00)	0.532
AFP (ng/ml)	70.70 (9.18, 973.35)	55.00 (6.80, 572.00)	0.689
Diameter (mm)			0.650
<50	35 (64.8%)	84 (68.3%)
≥50	19 (35.2%)	39 (31.7%)
Number			0.117
Single	27 (50.0%)	46 (37.4%)
Multiple	27 (50.0%)	77 (62.6%)
Differentiation			0.588
High	7 (13.0%)	10 (8.1%)
Medium	39 (72.2%)	92 (74.8%)
Low	8 (14.8%)	21 (17.1%)
T stage			0.954
T1–T2	38 (70.4%)	86 (69.9%)
T3–T4	16 (29.6%)	37 (30.1%)
SUVmax	4.97 (3.63, 6.21)	4.67 (3.43, 7.44)	0.909
SUVmean	2.89 (2.29, 3.32)	2.95 (2.34, 4.07)	0.426
SUVpeak	4.09 (3.02, 4.91)	4.02 (3.08, 6.16)	0.385
TLG (g/ml×cm³)	137.00 (67.75, 364.75)	211.00 (60.00, 618.00)	0.277
MTV (cm³)	60.95 (20.00, 157.95)	76.00 (28.00, 213.70)	0.163
COV	0.20 (0.15, 0.25)	0.19 (0.14, 0.26)	0.720
HI	33.55 (10.05, 79.90)	29.10 (9.40, 115.15)	0.600
TLR	1.56 (1.27, 1.97)	1.55 (1.16, 2.36)	0.803

To ensure generalizability, the model was rigorously validated on a prospective partitioned cohort (30% of the total data, n = 54). The baseline features showed no significant difference between the training set and the validation set (all P > 0.05). The column chart maintained robust discriminability in the validation queue (AUC = 0.815, 95% CI: 0.754–0.876), confirming its stability beyond the derived dataset.

AFP, alpha-fetoprotein; CI, confidence interval; COV, coefficient of variation; HI, heterogeneity index; MTV, metabolic tumor volume; SUV_max, maximum standardized uptake value; SUV_mean, mean standardized uptake value; SUV_peak, peak standardized uptake value; TLG, total lesion glycolysis; TLR, tumor-to-liver ratio.

The column chart showed consistent performance in the independent validation set, and the Hosmer-Lemeshow test (P = 0.1834, Fig. 4) confirmed good calibration and consistency between predicted and observed MVI risks. In addition, the discriminative power is still strong, with an AUC of 0.815 (95% CI: 0.754–0.876), which is comparable to the performance of the training set (AUC = 0.842) (Table 6, Fig. 5).

Fig. 4.

The calibration curve (Fig. 2) and Hosmer-Lemeshow test (P = 0.183) indicated good agreement between predicted and observed MVI probabilities in the validation cohort (mean absolute error = 0.03), supporting model reliability. MVI, microvascular invasion.

Table 6.

ROC curve for predicting microvascular invasion in hepatocellular carcinoma

Variable	AUC (95% CI)	P value	Sensitivity	Specificity	Threshold
Pivka-II (mAU/ml)	0.685 (0.591, 0.779)	<0.001	0.412	0.882	1769.000
TLG	0.709 (0.616, 0.802)	<0.001	0.618	0.782	246.500
TLR	0.729 (0.639, 0.820)	<0.001	0.662	0.745	1.556
HI	0.753 (0.668, 0.839)	<0.001	0.647	0.818	41.575
Combined	0.842 (0.773, 0.911)	<0.001	0.809	0.764	0.40038

HI, heterogeneity index; TLG, total lesion glycolysis; TLR, tumor-to-liver ratio.

Fig. 5.

ROC curve for predicting microvascular invasion in hepatocellular carcinoma.

Discussion

This study analyzed clinical and PET/CT imaging data from 177 HCC liver transplantation patients. Univariate analysis showed significant differences in PIVKA II, SUVmax, SUVmean, SUVpeak, TLG, TLR, COV, and HI between the two groups (P < 0.001). Multivariate analysis showed that PIVKA II, TLG, HI, and TLR are independent predictors of MVI. A nomogram prediction model was constructed based on logistic regression analysis. ROC showed that the combined prediction model had an AUC value of 0.815, significantly superior to single parameters, indicating that the integration of multimodal parameters can significantly improve the accuracy of preoperative MVI prediction.

Despite significant advances in the diagnosis and treatment of HCC in recent years, the risk of local recurrence and metastasis after liver transplantation remains significant. Therefore, screening patients who develop early HCC recurrence after liver transplantation and providing them with timely and personalized treatment may bring long-term benefits [17]. Multiple studies have shown that MVI is an important factor in predicting postoperative recurrence and is associated with poor prognosis in patients [7]. ¹⁸F-FDG PET/CT is an important molecular imaging examination method in clinical practice. Fluorine-18-fluorodeoxyglucose (fluor-18-deoxy-glucose, ¹⁸F-FDG) is currently the most commonly used PET imaging agent in clinical practice. It appears as an abnormal accumulation of the imaging agent in imaging studies, with vigorous glucose uptake by tumor cells, which is used to distinguish between benign and malignant tumors and even to determine the degree of differentiation [18–20].

SUVmax is the most commonly used semi-quantitative parameter in PET/CT and is also an important factor in assessing tumor prognosis and recurrence [21–23]. Previous studies have indicated that ¹⁸F-FDG PET/CT plays an important role in the preoperative prediction of MVI in HCC [15,16]. In this study, the differences in SUVmax, SUVmean, and SUVpeak between the two groups were statistically significant (all P < 0.001), consistent with the findings of Jiang and Hyun et al. [15,24]. Some studies consider TLR and TLG as the most commonly used metabolic parameters for evaluating PET images because they correlate better with the doubling time of HCC and reflect the metabolic activity of lesions more accurately than SUVmax [19]. The effectiveness of using TLR and TLG metabolic parameters to predict MVI in HCC patients has been confirmed by multiple studies [25–27]. Our study also included this in the analysis, indicating that TLR and TLG showed statistically significant differences between the two groups (P < 0.001). Univariate analysis confirmed significant associations between elevated SUVmax, SUVmean, SUVpeak, TLR, TLG, and MVI status (all P < 0.001), consistent with prior studies [15,20,24]. However, in multivariate analysis, only TLG, HI, TLR, and PIVKA-II emerged as independent predictors. This suggests that TLG (integrating metabolic activity and tumor volume) and HI (reflecting metabolic heterogeneity) may subsume or correlate with the metabolic information conveyed by SUV parameters, explaining their lack of independent contribution when stronger predictors are considered. HI demonstrated the highest specificity (0.831), indicating its utility in ruling out MVI-negative cases, while TLR offered optimal sensitivity (0.760).

The moderate collinearity observed between TLG/HI and SUV parameters (variance inflation factor > 2) supports the hypothesis that these composite indicators capture a wider range of tumor features than SUV values alone. This highlights the superiority of the combined column chart model (AUC = 0.815), which utilizes complementary predictive factors such as metabolic load (TLG), heterogeneity (HI), tumor liver contrast (TLR), and serological biomarkers (PIVKA-II) to achieve much higher accuracy than any single parameter (HI AUC = 0.754; TLR AUC = 0.744; P < 0.001).

Our findings align with Kim and Kim‘s [28] meta-analysis, which reported a limited standalone value of conventional PET parameters (e.g. SUVmax) for MVI prediction. By innovatively integrating HI with TLG and clinical biomarkers, our model overcomes this limitation, achieving clinically relevant accuracy (AUC > 0.8) suitable for risk stratification. This aligns with our observation that conventional parameters (SUVmax) failed as independent predictors in multivariate analysis. To overcome this limitation, we innovatively incorporated tumor heterogeneity indices, which is a dimension largely unexplored for MVI prediction in HCC. While heterogeneity parameters like COV and HI have prognostic value in other cancers [11–13], only two recent HCC studies [20,27,28] preliminarily linked HI to MVI. Our methodology adapted established approaches: COV (SUVsd/SUVmean [12], and HI) slope of MTV regression across SUV thresholds [13,14]. HI was calculated using a slope-based method validated in oncology heterogeneity studies [13,14,27].

HI demonstrated strong independent predictive power (OR = 1.022, P < 0.001), corroborating our study findings that HI outperformed traditional PET metrics. Although COV was insignificant in multivariate analysis, its synergy with HI warrants future investigation, especially given emerging evidence that combined heterogeneity markers improve risk stratification.

Our study also extends this by integrating HI with key metabolic parameters (TLG, TLR) and serum PIVKA-II to create a clinically applicable nomogram for liver transplant candidates. This multimodal approach achieved superior predictive accuracy (AUC = 0.815) compared to single parameters, addressing a critical need for individualized MVI risk stratification before transplantation.

In addition to the ¹⁸F-PET/CT metabolic parameters, early studies have found that serum AFP, serum PIVKA II, tumor number, and size are important independent indicators of MVI. These variables are closely associated with MVI and HCC recurrence [29–32]. However, in this study, only PIVKA II was found to be an important and independent factor for MVI, while the other parameters did not show differences, which is inconsistent with previous studies. The main reasons for this discrepancy may be the different sample sizes and a certain degree of selection bias. In a previous study, Yoh et al. [25] also used serum tumor markers AFP and ¹⁸F-FDG PET/CT metabolic parameter TLR to establish a multivariate linear regression model to predict MVI. Nomograms that incorporate various risk factors have been used to predict patient prognosis and treatment response. Based on our research findings, we developed a personalized nomogram model that integrates ¹⁸F-FDG PET/CT metabolic parameters, tumor heterogeneity parameters, and laboratory test indicators to more accurately predict MVI in HCC patients. The ROC results showed that HI (AUC = 0.754, 95% CI: 0.684–0.825), TLR (AUC = 0.744, 95% CI: 0.668–0.820), PIVKA II (AUC = 0.710, 95% CI: 0.643–0.786), and TLG (AUC = 0.716, 95% CI: 0.641–0.791), with a combined variable AUC reaching 0.815 (95% CI: 0.754–0.876), significantly improved the AUC value for predicting MVI compared to individual predictions, indicating that this model can effectively identify the risk of MVI. In addition, Wu et al. [27] specifically demonstrated the prognostic value of utilizing metabolic parameters derived from the PET component and heterogeneity metrics calculated from PET data coregistered with CT in HCC patients after liver transplantation. Bauschke et al. [33] found that TLR is an independent predictor of the 10-year cumulative recurrence rate. The study by Detry et al. [34] showed that the overall survival rates at 3 and 5 years after LT were 80.7% and 67.4%, respectively, while the recurrence-free survival rates were 70.6 and 67.4%. The nomogram model for predicting MVI can assist doctors in selecting the optimal treatment plan for HCC patients, improve treatment outcomes and clinical care, and reduce the overall recurrence rate after treatment.

This study has certain limitations, as it is a retrospective study with a relatively small sample size, which may lead to selection bias. Additionally, due to the short follow-up period, survival analysis is lacking. Furthermore, we will continue to follow up on the prognosis of these patients and plan to increase the sample size for a comprehensive survival analysis in the future. Future research needs to conduct larger-scale prospective studies that combine various imaging techniques and clinical indicators to establish a more comprehensive and accurate preoperative assessment model. Additionally, with the development of artificial intelligence technology, future research can explore the use of deep learning and other techniques for automated analysis and prediction of imaging data to improve diagnostic efficiency and accuracy. While this study adhered to TRIPOD guidelines in reporting the prediction model (Supplementary Table 1, Supplemental digital content 1, https://links.lww.com/NMC/A337), several limitations should be acknowledged. The retrospective design may introduce selection bias, although cohort partitioning was employed to mitigate overfitting. Additionally, external validation remains pending, necessitating future multicenter studies to assess the model’s generalizability. Finally, full compliance with TRIPOD guidelines requires prospective validation, which is planned as part of our ongoing research. What’s more, although propensity score matching (PSM) has shown significant value in this study, due to the limited sample size of the current cohort (N = 177), we mainly rely on internal validation methods to ensure the rigor of the results. Future research can expand the sample size through multicenter collaboration, and on this basis, apply PSM methods, which can not only more effectively control confounding bias, but also further verify the robustness of conclusions through subgroup analysis and sensitivity testing. This improved research design will help improve the generalization ability of the results and provide more reliable high-level evidence for clinical practice.

Acknowledgements

We would like to thank all the participants of this study and others who supported this study, including the funders.

Supported by the Major Project of Hangzhou Health and Family Planning Science and Technology Plan (ZD20220053).

The study protocol complied with the principles of the Declaration of Helsinki. This study was approved by the Institutional Review Board of Hangzhou Shulan Hospital (KY2022011). Confirming it as a retrospective study that meets the ethical requirements for waiver of informed consent. All data were collected in an electronic medical record system. All methods were performed in accordance with the relevant guidelines and regulations. There have been several researchers to check the accuracy and completeness of the data to avoid information bias. In addition, researchers anonymize or deidentify patient data to reduce the risk of data breaches.

J.W. and X.K. wrote the main manuscript text and prepared all figures. All authors reviewed the manuscript.

Conflicts of interest

There are no conflicts of interest.