Abstract
Objective
The aim of our study was to develop and validate a machine learning model for the preoperative identification of microvascular invasion (MVI) in patients with hepatocellular carcinoma (HCC).
Methods
This retrospective multicenter study was conducted in China. Patients with HCC from June 2017 to December 2024 were enrolled. Database was divided into training and internal validation sets randomly. Least absolute shrinkage and selection operator (LASSO) regression was employed for feature selection. Four machine learning algorithms were compared for MVI prediction. Model performance was evaluated using the area under the receiver operating characteristic (AUC), accuracy, sensitivity, specificity, precision, Youden's index, and F1 score. Finally, the machine learning model with the best performance was selected as our final model while using it for an independent external validation set. The SHapley Additive exPlanations (SHAP) diagram was utilized to elucidate the variable importance within the model, culminating in the amalgamation of the above metrics to discern the most succinct features.
Results
The study finally enrolled 496 patients, comprising 229 MVI-positive and 267 MVI-negative cases. A total of 42 patients with HCC were collected in the independent external validation center, of which 18 were MVI-positive. LASSO regression showed that AFP, tumor size, peripheral enhancement, mosaic architecture and washout start time were the significant predictors. Among the four models, the Gradient Boosting Machine (GBM) model showed the best performance in the internal validation set, with an AUC of 0.829. In the independent external validation set, the GBM model demonstrated an AUC of 0.812.
Conclusion
The machine learning model shows promising efficacy in preoperative MVI identification for HCC patients. This method has potential clinical applications and may help identify MVI preoperatively, potentially improving clinical outcomes.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12885-026-15828-3.
Keywords: Microvascular invasion, Hepatocellular carcinoma, Machine learning, Contrast-enhanced ultrasound, SonoVue
Introduction
Hepatocellular carcinoma (HCC) ranks as the sixth most common malignant tumor in the world and represents a primary liver cancer with high mortality rates [1, 2]. The 5-year survival rate of HCC is only 18%, largely attributable to the high recurrence rates after surgical treatment [3, 4]. As a significant threat to human health, HCC poses a major challenge in global healthcare field.
Microvascular invasion (MVI), defined histologically by the presence of cancer cell nests within the microvasculature of peritumoral liver tissue, constitutes a well-established independent risk factor for early recurrence, metastasis, and reduced long-term survival following curative treatment [5–7]. Tumor cells break through the vascular endothelium and enter the microvascular lumen, indicating the tumor’s aggressive and metastatic potential. Patients with MVI-positive HCC are significantly poorer prognosis compared to the MVI-negative patients. Currently, MVI diagnosis primarily relies on pathological examination. Due to the heterogeneity of tumors, the specimens obtained by preoperative biopsy are limited, MVI can only be diagnosed by histopathological examination of postoperative resected specimens [8], making MVI information unavailable to guide preoperative and intraoperative treatment decisions. Therefore, imaging examinations for predicting MVI are particularly important.
Imaging examinations including Contrast-enhanced ultrasound (CEUS), Contrast-enhanced CT, and Contrast-enhanced MRI have emerged as principal diagnostic techniques for HCC [6]. Among these, CEUS offers distinct advantages as a safe, radiation-free, and repeatable modality that enables real-time visualization of microcirculatory perfusion patterns in both the lesion and surrounding hepatic parenchyma, reflecting hemodynamic characteristics of the tumor and its microenvironment. As a result, CEUS has gradually become a first-line imaging method in the clinical diagnosis of HCC. The accurate preoperative prediction of MVI using CEUS may facilitate the development of personalized treatment strategies and optimization of preoperative management for HCC patients.
With the advancement of data science, machine learning (ML) algorithms, such as Random Forest (RF) and Adaptive Boosting (Ada), have shown significant potential in facilitating early intelligent diagnosis and prognosis analysis of diseases. For example, ML has been successfully applied to predict conditions including metabolic-associated fatty liver disease [9], liver fibrosis [10] and cardiovascular disease [11].
Based on a multicenter dataset, this study aims to develop and validate an interpretable ML model for the preoperative assessment of MVI in patients with HCC. The model will be constructed and tested using parameters including baseline clinical data, preoperative laboratory examination, conventional ultrasonography (US), and SonoVue-CEUS features. Furthermore, the predictive accuracy of different ML algorithms for MVI will be compared in the validation sets.
Materials and methods
Participants
We retrospectively collected patients with HCC who underwent CEUS before surgery from June 2017 to December 2024. The inclusion criteria were: (1) age ≥ 18 years; (2) no local or systemic treatment before CEUS; (3) obtained a definitive pathological diagnosis through surgical resection; (4) no allergies to any components of SonoVue. The exclusion criteria were: (1) age < 18 years; (2) patients with incomplete clinical data; (3) patients with incomplete pathological data; (4) patients with poor-quality ultrasound imaging.
Ultrasonographic and CEUS examination
Ultrasonographic examinations were performed using a convex array broadband probe (frequency 3.5–5.0 MHz). The probe was positioned to optimally visualize the target lesion and surrounding hepatic parenchyma. Then we observed and recorded the tumor characteristics. Once the probe was stabilized, it was switched to CEUS mode with a low mechanical index. A bolus of 1.0–2.4 ml of SonoVue (Bracco, Milan, Italy) was injected via the elbow vein, followed immediately by a 5 ml saline. The procedure was strictly timed and continuously recorded for 2 min. Dynamic images were then saved at 1-min intervals until 6 min after contrast administration. All imaging data were saved in DICOM format for subsequent analysis.
Data collection
This study constructed models based on demographic characteristics, clinical characteristics, laboratory information, conventional US and CEUS parameters. Features were included only if they had been tested at all research centers to ensure data consistency and comparability. A total of 39 variables were ultimately included, as detailed in Supplementary Table 1.
Both US and CEUS datasets were independently evaluated by two researchers, each possessing 5 years of specialized experience in abdominal radiology with particular expertise in hepatic imaging. In cases of discrepant interpretations, a third senior radiologist with 25 years of professional experience was consulted to achieve consensus.
MVI
MVI was defined as the presence of tumor cell nests within the microvasculatures of the adjacent liver parenchyma, which can be observed though a microscope [5, 12].
Statistical analyses
All statistical tests were performed using R (version 4.2.2), Python (version 3.10) and Free Statistics software (version 2.1), with pathological results as the gold standard. A two-sided test P value of less than 0.05 was considered significant.
To ensure analytical reliability and mitigate potential biases, patients with incomplete data were excluded from the final analytical dataset. Continuous variables with normal distribution were presented as means ± SD and analyzed by t-test. Non-normally distributed continuous variables were expressed as median and inter-quartile range and analyzed using the Mann–Whitney U test. Categorical factors were described as frequencies or proportions, and analyzed by χ2 test or Fisher’s exact test. Inter-observer variability analysis was performed using κ statistic, as detailed in the Supplementary Table 2. The agreement was classified as follows: poor for 0–0.2, fair for 0.2–0.4, moderate for 0.4–0.6, good for 0.6–0.8, and excellent for values greater than 0.8.
Data preprocessing
The dataset was first randomly partitioned into a training set (70%) and an internal validation set (30%). All subsequent steps were conducted exclusively on the training set, including feature selection and model training.
Feature selection
We used the least absolute shrinkage and selection operator (LASSO) regression for feature selection. fivefold cross-validation was used to determine the optimal value of λ. The optimal λ value and the corresponding features were determined based on the cross-validation performance in the training set.
Model training
Four commonly used supervised learning algorithms were used in this study, including Multi-layer Perceptron (MLP), Gradient Boosting Machine (GBM), RF and Ada. During model development, hyperparameter tuning was performed using fivefold cross-validation on the training set. The optimized models were subsequently evaluated on the validation sets.
Model evaluation
Model performance was assessed using multiple standard metrics in the validation sets. These included the area under the receiver operating characteristic (AUC) with its 95% confidence interval (95% CI), accuracy, sensitivity, specificity, precision, Youden's index, and F1‑score. Classification results were visualized through confusion matrices. Pairwise comparisons of model performance were conducted statistically using the DeLong test.
Model interpretability
SHapley Additive exPlanations (SHAP) was a technique for interpreting ML model predictions [13]. It quantified the importance of each characteristic in the model by calculating the contribution value. The magnitude of the Shapley value reflected the importance of the characteristic in the model's prediction. Our study utilized SHAP to improve the model's interpretability and transparency.
Ethical statement
All study procedures were conducted in accordance with the Declarations of Helsinki and Istanbul. The study was approved by the Ethics Committee of the First Hospital of Shanxi Medical University (No. KYLL 2023–132). Due to the retrospective nature of the study and the use of de-identified patient data, the requirement for informed consent was waived.
Results
Participants characteristics
The research flowchart was illustrated in Fig. 1. Following the application of inclusion and exclusion criteria, 496 patients with HCC were finally enrolled. The dataset was randomly divided into the training and internal validation set in a 7:3 ratio. An independent dataset was included as an external validation set. Ultimately, the final cohorts comprised 347 patients in the training set (mean age 56.53 ± 11.47 years; 285 males), 149 patients in the internal validation set (mean age 56.30 ± 11.04 years; 124 males), and 42 patients in the external validation set (mean age 57.69 ± 9.93 years; 33 males).The baseline features were summarized in Table 1.
Fig. 1.
Research flowchart
Table 1.
Characteristics of patients in three sets
| Variables | Training set (n = 347) | Internal validation set (n = 149) | External validation set (n = 42) |
|---|---|---|---|
| Age | 56.53 ± 11.47 | 56.30 ± 11.04 | 57.69 ± 9.93 |
| Gender | |||
| Female | 62 (17.87) | 25 (16.78) | 9 (21.43) |
| Male | 285 (82.13) | 124 (83.22) | 33 (78.57) |
| HBV | |||
| No | 81 (23.34) | 35 (23.49) | 15 (35.71) |
| Yes | 266 (76.66) | 114 (76.51) | 27 (64.29) |
| HCV | |||
| No | 332 (95.68) | 142 (95.3) | 40 (95.24) |
| Yes | 15 (4.32) | 7 (4.7) | 2 (4.76) |
| Alcohol | |||
| No | 302 (87.03) | 135 (90.6) | 34 (80.95) |
| Yes | 45 (12.97) | 14 (9.4) | 8 (19.05) |
| Hypertention | |||
| No | 263 (75.79) | 113 (75.84) | 26 (61.9) |
| Yes | 84 (24.21) | 36 (24.16) | 16 (38.1) |
| Diabetes | |||
| No | 289 (83.29) | 127 (85.23) | 32 (76.19) |
| Yes | 58 (16.71) | 22 (14.77) | 10 (23.81) |
| AFP, ng/mL | |||
| ≤ 20 | 153 (44.09) | 64 (42.95) | 16 (38.1) |
| 20–400 | 92 (26.51) | 40 (26.85) | 13 (30.95) |
| > 400 | 102 (29.39) | 45 (30.2) | 13 (30.95) |
| WBC,/L | |||
| ≤ 4*109 | 99 (28.53) | 35 (23.49) | 7 (16.67) |
| > 4*109 | 248 (71.47) | 114 (76.51) | 35 (83.33) |
| PLT,/L | |||
| ≤ 100*109 | 89 (25.65) | 34 (22.82) | 13 (30.95) |
| > 100*109 | 258 (74.35) | 115 (77.18) | 29 (69.05) |
| ALT, U/L | |||
| ≤ 44 | 237 (68.3) | 110 (73.83) | 25 (59.52) |
| > 44 | 110 (31.7) | 39 (26.17) | 17 (40.48) |
| AST, U/L | |||
| ≤ 40 | 216 (62.25) | 104 (69.8) | 26 (61.9) |
| > 40 | 131 (37.75) | 45 (30.2) | 16 (38.1) |
| Tumor size | 4.61 ± 2.61 | 4.65 ± 2.79 | 3.59 ± 2.46 |
| Location | |||
| Left lobe | 77 (22.19) | 45 (30.2) | 17 (40.48) |
| Right lobe | 270 (77.81) | 104 (69.8) | 25 (59.52) |
| Tumor shape | |||
| Elliptic | 236 (68.01) | 101 (67.79) | 30 (71.43) |
| Circle | 33 (9.51) | 13 (8.72) | 5 (11.9) |
| Irregular | 78 (22.48) | 35 (23.49) | 7 (16.67) |
| Tumor margin | |||
| Clear | 262 (75.5) | 113 (75.84) | 30 (71.43) |
| Unclear | 85 (24.5) | 36 (24.16) | 12 (28.57) |
| Tumor echogenicity | |||
| Hypo | 210 (60.52) | 85 (57.05) | 17 (40.48) |
| Iso | 20 (5.76) | 21 (14.09) | 3 (7.14) |
| Hyper | 73 (21.04) | 25 (16.78) | 6 (14.29) |
| Mixed | 44 (12.68) | 18 (12.08) | 16 (38.1) |
| Tumor echo distribution | |||
| Heterogeneous | 114 (32.85) | 46 (30.87) | 23 (54.76) |
| Homogeneous | 233 (67.15) | 103 (69.13) | 19 (45.24) |
| Blood flow signal | |||
| Negative | 142 (40.92) | 61 (40.94) | 25 (59.52) |
| Positive | 205 (59.08) | 88 (59.06) | 17 (40.48) |
| Liver bacground | |||
| Cirrhosis | 75 (21.61) | 27 (18.12) | 8 (19.05) |
| Increased echo | 200 (57.64) | 82 (55.03) | 16 (38.1) |
| Hepatic steatosis | 49 (14.12) | 26 (17.45) | 12 (28.57) |
| Normal | 23 (6.63) | 14 (9.4) | 6 (14.29) |
| Portal hypertension | |||
| No | 276 (79.54) | 120 (80.54) | 37 (88.1) |
| Yes | 71 (20.46) | 29 (19.46) | 5 (11.9) |
| Splenomegaly | |||
| No | 258 (74.35) | 121 (81.21) | 36 (85.71) |
| Yes | 89 (25.65) | 28 (18.79) | 6 (14.29) |
| Ascites | |||
| No | 329 (94.81) | 138 (92.62) | 40 (95.24) |
| Yes | 18 (5.19) | 11 (7.38) | 2 (4.76) |
| Feeding artery | |||
| Negative | 52 (14.99) | 21 (14.09) | 8 (19.05) |
| Positive | 295 (85.01) | 128 (85.91) | 34 (80.95) |
| Intratumoral vascularity | |||
| Negative | 17 (4.9) | 11 (7.38) | 2 (4.76) |
| Positive | 330 (95.1) | 138 (92.62) | 40 (95.24) |
| Envelope enhancement | |||
| Negative | 186 (53.6) | 90 (60.4) | 31 (73.81) |
| Positive | 161 (46.4) | 59 (39.6) | 11 (26.19) |
| AP-enhancement pattern | |||
| Overall | 266 (76.66) | 119 (79.87) | 33 (78.57) |
| Local | 37 (10.66) | 11 (7.38) | 5 (11.9) |
| Rim | 40 (11.53) | 18 (12.08) | 4 (9.52) |
| Centrifugal | 4 (1.15) | 1 (0.67) | 0 (0) |
| Circular enhancement | |||
| Negative | 281 (80.98) | 133 (89.26) | 37 (88.1) |
| Positive | 66 (19.02) | 16 (10.74) | 5 (11.9) |
| AP-enhanced level | |||
| Hypo | 1 (0.29) | 1 (0.67) | 1 (2.38) |
| Iso | 13 (3.75) | 6 (4.03) | 1 (2.38) |
| Hyper | 333 (95.97) | 142 (95.3) | 40 (95.24) |
| Peripheral enhancement | |||
| Negative | 264 (76.08) | 115 (77.18) | 35 (83.33) |
| Positive | 83 (23.92) | 34 (22.82) | 7 (16.67) |
| Mosaic | |||
| Negative | 246 (70.89) | 110 (73.83) | 34 (80.95) |
| Positive | 101 (29.11) | 39 (26.17) | 8 (19.05) |
| AP-enhanced homogeneity | |||
| Homogeneous | 167 (48.13) | 77 (51.68) | 27 (64.29) |
| Heterogeneous | 180 (51.87) | 72 (48.32) | 15 (35.71) |
| AP-enhanced margin | |||
| Well defined | 214 (61.67) | 81 (54.36) | 29 (69.05) |
| Poorly defined | 133 (38.33) | 68 (45.64) | 13 (30.95) |
| AP-enhanced shape | |||
| Regular | 171 (49.28) | 83 (55.7) | 26 (61.9) |
| Irregular | 176 (50.72) | 66 (44.3) | 16 (38.1) |
| Necrotic | |||
| Negative | 200 (57.64) | 86 (57.72) | 30 (71.43) |
| Positive | 147 (42.36) | 63 (42.28) | 12 (28.57) |
| PVP-enhanced level | |||
| Hypo | 263 (75.79) | 119 (79.87) | 32 (76.19) |
| Hyper or iso | 84 (24.21) | 30 (20.13) | 10 (23.81) |
| DP-enhanced level | |||
| Hypo | 310 (89.34) | 137 (91.95) | 41 (97.62) |
| Hyper or iso | 37 (10.66) | 12 (8.05) | 1 (2.38) |
| DP-degree of washout | |||
| Obvious | 161 (46.4) | 68 (45.64) | 22 (52.38) |
| Mild/Moderate | 171 (49.28) | 77 (51.68) | 20 (47.62) |
| None | 15 (4.32) | 4 (2.68) | 0 (0) |
| Washout start time | |||
| ≤ 60 | 189 (54.47) | 92 (61.74) | 25 (59.52) |
| > 60 | 158 (45.53) | 57 (38.26) | 17 (40.48) |
| MVI | |||
| Negative | 187 (53.89) | 80 (53.69) | 24 (57.14) |
| Positive | 160 (46.11) | 69 (46.31) | 18 (42.86) |
Feature selection
As shown in Fig. 2, when λ.1se was chosen as the optimal λ value, five variables would be included in the model, including AFP, tumor size, peripheral enhancement, mosaic architecture and washout start time.
Fig. 2.
Features election using LASSO. A Log (lambda) value of 39 features in the LASSO model. A coefficient profile plot was produced against a log (lambda) sequence. B Cross-validation curves (fivefold cross validation). Partial likelihood deviation (binomial deviation) curves and logarithmic (lambda) curves are plotted. Minimum standard and 1-SE of the minimum standard are used to draw a vertical dashed line at the optimal value. Optimal lambda produces 5 nonzero coefficients. LASSO: least absolute shrinkage and selection operator
Model evaluation
Comprehensive model evaluation was performed using multiple approaches. The receiver operating characteristic (ROC), precision-recall curve (PRC), calibration curves and decision curve analysis (DCA) were shown in Fig. 3, with corresponding confusion matrices displayed in Fig. 4. Evaluation metrics of the predictive ability in the training and validation sets were summarized in Table 2. In the internal validation set, the GBM model shown superior discriminative ability, with the highest AUC of 0.829 (95%CI: 0.764–0.893). In the independent external validation set, the GBM model shown an AUC of 0.812 (95%CI: 0.659–0.965), as shown in Fig. 5.
Fig. 3.
Performance of machine learning models to predict MVI. The ROC (A), PRC (B), calibration curve (C) and the decision curve analysis (D) of the machine learning model in the training set. The ROC (E), PRC (F), calibration curve (G) and the decision curve analysis (H) of the machine learning model in the internal validation set
Fig. 4.
Confusion Matrices of machine learning models. Confusion matrices for GBM (A), MLP (B), RF (C) and Ada (D) in the training set. Confusion matrices for GBM (E), MLP (F), RF (G) and Ada (H) in the internal validation set
Table 2.
Comparison of the predictive ability of four models
| Model | AUC | 95%CI | Accuracy | Sensitivity | Specificity | Precision | Youden | F1 |
|---|---|---|---|---|---|---|---|---|
| Training set | ||||||||
| GBM | 0.853 | 0.811–0.895 | 0.793 | 0.762 | 0.818 | 0.782 | 0.580 | 0.772 |
| MLP | 0.864 | 0.824–0.904 | 0.775 | 0.706 | 0.834 | 0.785 | 0.540 | 0.743 |
| RF | 0.827 | 0.783–0.871 | 0.772 | 0.762 | 0.781 | 0.748 | 0.543 | 0.755 |
| Ada | 0.825 | 0.781–0.870 | 0.778 | 0.850 | 0.717 | 0.720 | 0.567 | 0.779 |
| Internal validation set | ||||||||
| GBM | 0.829 | 0.764–0.893 | 0.718 | 0.710 | 0.725 | 0.690 | 0.435 | 0.700 |
| MLP | 0.801 | 0.729–0.874 | 0.745 | 0.652 | 0.825 | 0.763 | 0.477 | 0.703 |
| RF | 0.822 | 0.756–0.888 | 0.732 | 0.725 | 0.738 | 0.704 | 0.462 | 0.714 |
| Ada | 0.823 | 0.757–0.889 | 0.711 | 0.855 | 0.588 | 0.641 | 0.443 | 0.733 |
| External validation set | ||||||||
| GBM | 0.812 | 0.659–0.965 | 0.762 | 0.722 | 0.792 | 0.722 | 0.514 | 0.723 |
| MLP | 0.829 | 0.698–0.959 | 0.810 | 0.667 | 0.817 | 0.857 | 0.484 | 0.750 |
| RF | 0.807 | 0.662–0.951 | 0.738 | 0.721 | 0.750 | 0.684 | 0.471 | 0.703 |
| Ada | 0.809 | 0.669–0.949 | 0.810 | 0.722 | 0.875 | 0.812 | 0.597 | 0.765 |
GBM Gradient Boosting Machine, MLP Multi-layer Perceptron, RF Random Forest, Ada Adaptive Boosting
Fig. 5.
Performance of GBM model to predict MVI in the external validation set
Supplementary Table 3 presented the results of pairwise statistical comparisons of the AUCs among different models using the DeLong test. The GBM model demonstrated a statistically significant superior performance compared to the MLP model (P = 0.037).
Model interpretability
Figure 6 displayed the bar chart of SHAP values, which quantitatively visualized the overall importance of each feature. “Washout start time” exhibited the highest SHAP value of + 0.86, indicating it was the most influential feature.
Fig. 6.
Model explanation by the SHAP method. A Beeswarm plot of the GBM Model. Generate SHAP values for each variable and reveal its relationship with MVI. B SHAP feature importance for GBM model
Discussion
Currently, MVI is the most important pathological feature to guide the prevention and treatment of HCC recurrence, which primarily relies on histopathological examination of postoperative resected specimen for diagnosis, resulting in certain delays. Curative therapies such as surgery or liver transplantation must be decided preoperatively, which may lead to potential issues such as low-risk patients undergoing excessive treatment due to expanded liver resection, or high-risk patients experiencing early recurrence due to positive local margins postoperatively. It highlights the critical necessity and importance of preoperative prediction of MVI. This study was designed to develop and validate an explainable ML model based on preoperative US and CEUS images for the noninvasive preoperative assessment of MVI in patients with HCC, providing a more accurate predictive tool for clinical practice, improving the ability to accurately assess MVI risk preoperatively, and offering new reference criteria for personalized treatment and risk assessment for patients.
Our results showed that MVI-positive cases had higher AFP levels and larger tumor size; CEUS showed peripheral enhancement, mosaic architecture, and early washout. AFP is one of the important serum markers for the diagnosis of HCC as well as the key indicator for monitoring the efficacy of HCC treatment. It has been shown that AFP was associated with the occurrence of MVI [14, 15]. Although current studies differed in the definition of AFP cutoff values, they all confirmed that elevated AFP levels were positively correlated with MVI risk. Tumor size is also closely related to the presence of MVI. Several studies have confirmed that tumor size was an independent risk factor for predicting the occurrence of MVI, and the risk of MVI increased with tumor size [16–19].
During the process of metastasis, tumor cells initially invade the nearby microvasculature, then further spread through the microvasculature surrounding the tumor. As the interface between tumor and normal tissues, the peritumoral tissue provides a more accurate reflection of tumor growth, invasion, and metastasis. Peritumoral enhancement was associated with the tumor's arterial supply and vascularization. During tumor invasive growth, invasion of surrounding tiny portal vein branches occurs, leading to MVI, which alters local hemodynamics within the liver tissue. Newly formed vessels compensate with increased perfusion, resulting in significant hyperenhancement of the peritumoral liver tissue during arterial phase (AP). Several studies confirmed that peritumoral enhancement was positively correlated with the risk of MVI [20–24].
Mosaic architecture is a marker of tumor heterogeneity and an auxilary feature in the LI-RADS. When tumor cells invade surrounding microvessels and disrupt vascular integrity, both structural damage and increased vascular wall fragility render intratumoral vessels more vulnerable to rupture and hemorrhage. Furthermore, vascular disruption reduces intratumoral perfusion, leading to tissue hypoxia and nutrient deprivation, which in turn causes localized ischemia and necrosis. These alterations in the tumor microenvironment promote the formation of numerous disorganized neovasculature while enhancing tumor cell invasiveness, creating conditions conducive to vascular invasion. These factors contribute to the development of the mosaic architecture. Mosaic architecture manifests as numerous randomly distributed internal nodules or cystic structures, separated by fibrous components, and exhibiting considerable heterogeneity in size, morphology and enhancement patterns [25, 26]. The enhanced areas correspond to viable tumor tissue, while non-enhanced areas indicate regions of necrosis, hemorrhage, or fibrosis [27]. Wang Y et al. found that the mosaic architecture on CEUS was an independent predictor of MVI [28]. Our study obtained similar results.
What’s more, early washout was highly correlated with MVI risk, washout was related to the microcirculatory perfusion of the lesion [29]. MVI-positive HCCs were typically poorly differentiated, with the hepatic artery as the primary feeding vessel. Additionally, due to the absence of hepatic sinusoidal structures and the presence of arteriovenous fistulas, contrast agents rapidly passed through tumor vessels, resulting in early and rapid washout. Yao et al. conducted a prospective multicenter study enrolling 211 HCC patients. The results showed that washout time (< 90 s) was significantly correlated with MVI [30].
We employed the SHAP explainer to visually explain the output of the GBM model, which demonstrated the highest diagnostic performance. GBM is a ML classifier widely applied to classification tasks. The SHAP beeswarm plot illustrates the influence range and distribution of feature's contribution to the model's output. The SHAP bar plot can be used to compare the global importance of features [13]. Our results showed that washout start time was the most critical feature for MVI, followed by mosaic architecture, tumor size, peripheral enhancement and AFP. Physicians can directly assess the relationship between a specific feature and an individual patient's MVI risk by observing the distribution pattern of that feature's values in relation to their corresponding SHAP values.
The results of our study confirm the feasibility of using imaging information for noninvasive assessment of MVI. The model developed and validated in this study is primarily based on clinical and US characteristics. Compared to models that rely on expensive and radiative MRI or CT, CEUS has the advantages of being safe, radiation-free, and reproducible. What’s more, CEUS provides real-time, dynamic visualization of microcirculatory perfusion of contrast agent, reflecting hemodynamic changes within and around the lesion, which is of vital importance for the precise diagnosis of HCC.
There are several challenges in this study. Firstly, it was a retrospective study. Our conclusions should be further validated in a large-scale, prospective, multicenter study in the future to assess their universality. What’s more, the parameters used in this study relied on physicians' professional interpretations, which may be subjective. In the future, we will use deep learning to automatically extract US imaging features to further improve the model's performance.
Conclusion
In this study, we applied four machine learning methods to develop and validate an ultrasound-based prediction model for preoperative evaluation of MVI in patients with HCC, applying multiple metrics to evaluate the models’ performance, and finally selecting the best predictive performance of the ML model, GBM. The model has the potential to provide a non-invasive diagnosis of MVI preoperatively clinicians that enables precise risk stratification, optimizes treatment decisions, and facilitates personalized patient management.
Supplementary Information
Supplementary Material 1: Supplementary Table 1. Variables Collected for Model Development.
Supplementary Material 2: Supplementary Table 2. Inter-observer variability analysis.
Supplementary Material 3: Supplementary Table 3. Comparison of model AUCs using the DeLong test.
Acknowledgements
Not applicable.
Abbreviations
- MVI
Microvascular invasion
- HCC
Hepatocellular carcinoma
- CEUS
Contrast-enhanced ultrasound
- ML
Machine learning
- RF
Random Forest
- Ada
Adaptive Boosting
- US
Ultrasound
- AFP
Alpha-fetoprotein
- WBC
White blood cells
- PLT
Platelets
- ALT
Alanine aminotransferase
- AST
Aspartate aminotransferase
- LASSO
Least absolute shrinkage and selection operator
- MLP
Multi-layer Perceptron
- GBM
Gradient Boosting Machine
- AUC
Area under the receiver operating characteristic
- CI
Confidence interval
- SHAP
SHapley Additive exPlanations
- ROC
Receiver operating characteristic
- PRC
Precision-recall curve
- DCA
Decision curve analysis
- AP
Arterial phase
- PVP
Portal venous phase
- DP
Delay phase
Authors’ contributions
ZQ wrote the manuscript; LLP designed the manuscript; PC, LP, SLL, LZL, WRN and FWW revised the manuscript. All authors read and approved the final manuscript.
Funding
This work was supported by the National Natural Science Foundation of China (82272026); Shanxi Scholarship Council of China (2022–193); Special project of scientific and technological cooperation and exchange in Shanxi Province (202304041101031).
Data availability
The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.
Declarations
Ethics approval and consent to participate
All study procedures were conducted in accordance with the Declarations of Helsinki and Istanbul. The study was approved by the Ethics Committee of the First Hospital of Shanxi Medical University (No. KYLL 2023–132). Due to the retrospective nature of the study and the use of de-identified patient data, the requirement for informed consent was waived.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Qian Zhang and Chuan Pang contributed equally to this work.
Contributor Information
Ping Liang, Email: liangping301@hotmail.com.
Liping Liu, Email: liuliping1600@sina.com.
References
- 1.Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74:229–63. [DOI] [PubMed] [Google Scholar]
- 2.Siegel RL, Miller KD, Wagle NS, et al. Cancer statistics, 2023. CA Cancer J Clin. 2023;73(1):17–48. [DOI] [PubMed] [Google Scholar]
- 3.Vogel A, Meyer T, Sapisochin G, et al. Hepatocellular carcinoma. Lancet. 2022;400(10360):1345–62. [DOI] [PubMed] [Google Scholar]
- 4.Yang J, Pan G, Guan L, et al. The burden of primary liver cancer caused by specific etiologies from 1990 to 2019 at the global, regional, and national levels. Cancer Med. 2022;11(5):1357–70. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Tang M, Zhang S, Yang M, et al. Infiltrative vessel co-optive growth pattern induced by IQGAP3 overexpression promotes microvascular invasion in hepatocellular carcinoma. Clin Cancer Res. 2024;30(10):2206–24. [DOI] [PubMed] [Google Scholar]
- 6.Sun J, Xia Y, Shen F, et al. Chinese expert consensus on the diagnosis and treatment of hepatocellular carcinoma with microvascular invasion (2024 edition). Hepatobiliary Surg Nutr. 2025;14(2):246–66. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Lee S, Kang TW, Song KD, et al. Effect of microvascular invasion risk on early recurrence of hepatocellular carcinoma after surgery and radiofrequency ablation. Ann Surg. 2021;273(3):564–71. [DOI] [PubMed] [Google Scholar]
- 8.Zhang ZH, Jiang C, Qiang ZY, et al. Role of microvascular invasion in early recurrence of hepatocellular carcinoma after liver resection: a literature review. Asian J Surg. 2024;47(5):2138–43. [DOI] [PubMed] [Google Scholar]
- 9.Deng J, Ji W, Liu H, et al. Development and validation of a machine learning-based framework for assessing metabolic-associated fatty liver disease risk. BMC Public Health. 2024;24(1):2545. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Fan R, Yu N, Li G, et al. Machine-learning model comprising five clinical indices and liver stiffness measurement can accurately identify MASLD-related liver fibrosis. Liver Int. 2024;44(3):749–59. [DOI] [PubMed] [Google Scholar]
- 11.Drożdż K, Nabrdalik K, Kwiendacz H, et al. Risk factors for cardiovascular disease in patients with metabolic-associated fatty liver disease: a machine learning approach. Cardiovasc Diabetol. 2022;21(1):240. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Roayaie S, Blume IN, Thung SN, et al. A system of classifying microvascular invasion to predict outcome after resection in patients with hepatocellular carcinoma. Gastroenterology. 2009;137(3):850–5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Hu J, Xu J, Li M, et al. Identification and validation of an explainable prediction model of acute kidney injury with prognostic implications in critically ill children: a prospective multicenter cohort study. EClinicalMedicine. 2024;68:102409. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Zhang L, Li M, Zhu J, et al. The value of quantitative MR elastography-based stiffness for assessing the microvascular invasion grade in hepatocellular carcinoma. Eur Radiol. 2023;33(6):4103–14. [DOI] [PubMed] [Google Scholar]
- 15.Jin J, Niu X, Zou L, et al. AFP mRNA level in enriched circulating tumor cells from hepatocellular carcinoma patient blood samples is a pivotal predictive marker for metastasis. Cancer Lett. 2016;378(1):33–7. [DOI] [PubMed] [Google Scholar]
- 16.Li Z, Lei Z, Xia Y, et al. Association of preoperative antiviral treatment with incidences of microvascular invasion and early tumor recurrence in hepatitis B virus-related hepatocellular carcinoma. JAMA Surg. 2018;153(10):e182721. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Xu X, Zhang HL, Liu QP, et al. Radiomic analysis of contrast-enhanced CT predicts microvascular invasion and outcome in hepatocellular carcinoma. J Hepatol. 2019;70(6):1133–44. [DOI] [PubMed] [Google Scholar]
- 18.He YZ, He K, Huang RQ, et al. Preoperative evaluation and prediction of clinical scores for hepatocellular carcinoma microvascular invasion: a single-center retrospective analysis. Ann Hepatol. 2020;19(6):654–61. [DOI] [PubMed] [Google Scholar]
- 19.Zhu Z, Wu K, Lu J, et al. Gd-EOB-DTPA-enhanced MRI radiomics and deep learning models to predict microvascular invasion in hepatocellular carcinoma: a multicenter study. BMC Med Imaging. 2025;25(1):105. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Min JH, Kim YK, Choi SY, et al. Differentiation between cholangiocarcinoma and hepatocellular carcinoma with target sign on diffusion-weighted imaging and hepatobiliary phase gadoxetic acid-enhanced MR imaging: classification tree analysis applying capsule and septum. Eur J Radiol. 2017;92:1–10. [DOI] [PubMed] [Google Scholar]
- 21.Lee S, Kim SH, Lee JE, et al. Preoperative gadoxetic acid-enhanced MRI for predicting microvascular invasion in patients with single hepatocellular carcinoma. J Hepatol. 2017;67(3):526–34. [DOI] [PubMed] [Google Scholar]
- 22.Renzulli M, Brocchi S, Cucchetti A, et al. Can current preoperative imaging be used to detect microvascular invasion of hepatocellular carcinoma? Radiology. 2016;279(2):432–42. [DOI] [PubMed] [Google Scholar]
- 23.Zheng R, Zhang X, Liu B, et al. Comparison of non-radiomics imaging features and radiomics models based on contrast-enhanced ultrasound and Gd-EOB-DTPA-enhanced MRI for predicting microvascular invasion in hepatocellular carcinoma within 5 cm. Eur Radiol. 2023;33(9):6462–72. [DOI] [PubMed] [Google Scholar]
- 24.Zhang HD, Li XM, Zhang YH, et al. Evaluation of preoperative microvascular invasion in hepatocellular carcinoma through multidimensional parameter combination modeling based on Gd-EOB-DTPA MRI. J Clin Transl Hepatol. 2023;11(2):350–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Bartolotta TV, Terranova MC, Gagliardo C, et al. CEUS LI-RADS: a pictorial review. Insights Imaging. 2020;11(1):9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Caraiani C, Boca B, Bura V, et al. CT/MRI LI-RADS v2018 vs. CEUS LI-RADS v2017-can things be put together? Biology Basel. 2021;10(5):412. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Sheng Y, Wang Q, Liu HF, et al. Preoperative nomogram incorporating clinical factors, serological markers and LI-RADS MRI features to predict early recurrence of hepatocellular carcinoma treated with transarterial chemoembolization. Acad Radiol. 2023;30(7):1288–97. [DOI] [PubMed] [Google Scholar]
- 28.Wang Y, Chai S, Cai W, et al. Prognostic and pathological implications of contrast-enhanced ultrasound features in hepatocellular carcinoma. J Cancer Res Ther. 2023;19(4):1040–7. [DOI] [PubMed] [Google Scholar]
- 29.Zhang Q, Liang X, Zhang Y, et al. A review of contrast-enhanced ultrasound using SonoVue® and Sonazoid™ in non-hepatic organs. Eur J Radiol. 2023;167:111060. [DOI] [PubMed] [Google Scholar]
- 30.Yao J, Li K, Yang H, et al. Analysis of Sonazoid contrast-enhanced ultrasound for predicting the risk of microvascular invasion in hepatocellular carcinoma: a prospective multicenter study. Eur Radiol. 2023;33(10):7066–76. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Supplementary Material 1: Supplementary Table 1. Variables Collected for Model Development.
Supplementary Material 2: Supplementary Table 2. Inter-observer variability analysis.
Supplementary Material 3: Supplementary Table 3. Comparison of model AUCs using the DeLong test.
Data Availability Statement
The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.