Vascular Invasion Prediction Via a Multimodal Model Supports Re-staging of Early-stage Hepatocellular Carcinoma

1. Study rationale and scope
2. Individualized treatment decisions for early-stage HCC remain unresolved. Surgical resection, liver transplantation, and local ablation are the first-line treatments for early-stage HCC. Treatment selection using traditional indicators such as tumor burden, location, liver function, and technical feasibility can be unreliable. Significant interpatient heterogeneity results in many patients failing to achieve optimal benefit from the selected treatment, leading to low survival rates, high complication rates, and high medical costs. Incorporating tumor biological features into treatment decisions may mitigate the impact of interpatient heterogeneity. However, due to technical limitations, the currently proposed molecular subtyping and pathomolecular classifications, which aim to guide biology-based decisions, are not recommended by recent clinical guidelines for early-stage HCC treatment planning.
3. MVI predictiion has the potential to guide individualized treatment decisions. Microvascular invasion (MVI), a hallmark of HCC invasiveness, is present in 20.0% to 45.9% of resected early-stage HCC specimens and strongly indicates a high recurrence risk and the need for adjuvant therapy. Current studies have highlighted the critical and potential value of preoperative MVI status to guide initial treatment selection. In brief, for patients with MVI, more aggressive treatments may effectively reduce recurrence, while for those without MVI, less aggressive therapies can ensure outcomes while being safer and more cost-effective. These findings underscore the critical role of MVI in guiding treatment selection. However, since MVI can only be diagnosed through postoperative histopathology, its presence or absence preoperatively requires prediction. Moreover, current studies suggest that preoperative MVI prediction has the potential to guide not only the choice among resection, transplantation, and ablation but also specific technical decisions, such as resection type, resection or ablation margin width, and the use of neoadjuvant therapy. It also prompts a reconsideration of current transplantation criteria.
4. Integration of radiomics and liquid biopsy may improve the accuracy of MVI prediction.

Numerous studies have utilized clinical characteristics, tumor features from imaging studies, laboratory tests including tumor markers (e.g., α-fetoprotein [AFP], α-fetoprotein lens culinaris agglutinin-reactive fraction 3 [AFP-L3], prothrombin induced by vitamin K absence-II [PIVKA-II]), and multifactorial clinical models for MVI prediction. However, the sensitivity and specificity of these methods remain insufficient, limiting their widespread clinical application.

Radiomics, which encompasses techniques utilizing contrast-enhanced CT, MRI, ultrasound, and PET-CT, has demonstrated superior performance, with MRI-based radiomics being particularly notable. This advantage derives from MRI's excellent soft-tissue contrast and multiparametric imaging capabilities. These capabilities are further enhanced by machine learning (ML) and deep learning (DL) technologies, which capture more subtle tumor features, thereby improving the sensitivity and robustness of predictions. However, despite numerous successful reports, the clinical translation of radiomics faces challenges. Beyond technical factors, which are expected to be overcome with technological advancements, a limitation in tumor biology exists: radiomics-based prediction can be suboptimal when aggressive tumor biology has not yet manifested as discernible changes on imaging-a key limitation in predicting MVI for early-stage HCC.

Certain genomic alterations detected in HCC tumor tissues have the potential to predict MVI. However, current guidelines do not recommend preoperative biopsy for resectable HCC due to the risk of tumor needle tract dissemination. Additionally, intratumoral heterogeneity significantly affects testing accuracy.

Liquid biopsy, with its ability to detect minimal residual disease (MRD), may enable accurate MVI prediction. However, the application of current liquid biopsy methods is limited by low sensitivity, technical complexity, and an inability to fully characterize tumor heterogeneity. Circulating cell-free DNA (cfDNA) containing cancer-specific information can more accurately characterize tumor-invasive capability. Thus, cfDNA testing represents a novel approach with the potential to accurately estimate MVI risk, although supporting data remain scarce.

Furthermore, combining cfDNA analysis with radiomics may complementarily capture both the biological behavior and structural features of MVI, thereby improving predictive performance. However, this integrated approach remains unexplored.

(4) The strategy of predicting MVI to guide treatment decisions is only applicable to very early and early-stage HCC (BCLC stage 0 or A).

The treatment strategies for intermediate- and advanced-stage HCC differ from those for early-stage HCC, and the biomarkers guiding treatment decisions for these stages are different from or not limited to MVI.

1. Study design
2. Overall design: This is a large-scale, prospective, multicenter cohort study. This study enrolls resected early-stage HCC patients, aiming to explore the feasibility of optimizing surgical decisions through preoperative multimodal MVI prediction.
3. Development of the multimodal MVI-predictive model: The multimodal framework integrates predictions from four individual models: (1) a clinical model developed based on clinicopathological variables using the eXtreme Gradient Boosting (XGBoost) algorithm (n=3,000); (2) an MRI-based radiomics model, trained on the same cohort with contrast-enhanced MRI data; (3) an MRI-based deep learning model, trained on the same imaging cohort to extract high-level features; and (4) a cfDNA mutation model, trained on patients with cfDNA signatures predictive of MVI (n=650). The four component models will generate patient-level predicted probabilities of MVI. Within the training data, out-of-fold predictions will be used to estimate probability-calibration functions, with Platt scaling prespecified as the primary calibration method and isotonic regression evaluated in sensitivity analyses. The calibrated probabilities will be clipped to avoid boundary values of 0 and 1 and transformed to logits before being entered into an L2-regularized logistic stacking model. The fusion coefficients of the final multimodal system will be estimated exclusively from out-of-fold predictions. After all preprocessing, calibration, fusion and threshold-selection procedures have been locked, patients with a predicted MVI probability at or above the cutoff determined by the maximum Youden index will be classified as high risk, whereas those below the cutoff will be classified as low risk.
4. Judgment of prognostic stratification for the multimodal model: The early-stage HCC patients will be re-staged into the MVI low-risk and high-risk substages based on the optimal cutoff value of the multimodal model. The differences in 5-year OS, 5-year RFS, and recurrence patterns between patients in the two substages will be compared to evaluate the accuracy of this re-staging in stratifying clinical prognosis.
5. Evaluation of clinical utility for the multimodal model: The clinical value of the re-staging will be assessed in the clinical scenario of selecting the appropriate resection margin width. Patients in both the MVI high-risk and low-risk substages will be divided into two subgroups: wide-margin versus narrow-margin resection. The 5-year OS, 5-year RFS, recurrence patterns will be compared between these two subgroups. These results will indicate whether the re-staging system can accurately guide the surgical decisions in early-stage HCC.

This study has been approved by the Institutional Ethics Committee of each center, and informed consents have been obtained from all patients for their tissues or blood samples and clinical data to be used for research purposes.

1. Patients and enrollment (1) Inclusion criteria
2. Age between 18 and 75 years; 2) Histopathologically confirmed HCC; 3) Very early and early-stage HCC meeting BCLC stage 0 or A; 4) Child-Pugh class A or B liver function; 5) Undergoing curative-intent resection, defined as complete tumor removal with tumor-free resection margins, without major vascular invasion or extrahepatic distant metastasis; 6) No preoperative anti-cancer treatment prior to surgery; 7) Availability of preoperative contrast-enhanced MRI data eligible for imaging tumor feature assessment; 8) Availability of preoperative peripheral blood sample for cfDNA sequencing (only for patients with cfDNA seq mutation analysis); 9) Availability of complete clinicopathological and follow-up data; 10) Written informed consent for research use of clinicopathological data and biological samples. (2) Exclusion criteria
3. A history of other malignancies;
4. Residual tumors identified within 2 months after surgery;
5. Serious surgical complications per the Clavien-Dindo classification. (3) Patient enrollment China's National Center for Liver Cancer (NCLC)/Eastern Hepatobiliary Surgery Hospital (EHBH) maintains a prospectively collected database and biobank (supported by the State Key Project on Infectious Diseases [2018ZX10723-204] and the Key Project on Clinical Cohort Development of Shanghai [SHSLCZDZK02402]), which includes consecutive cases of surgically resected liver cancer between January 2015 and April 2026. We will identify study candidates from this database. A total of 3,000 eligible patients are planned to be enrolled. Their preoperative MRI data will be utilized to establish both radiomics and deep learning MVI-predictive models.

We have also established a prospective cohort of 650 patients from four Chinese tertiary medical centers: Zhongda Hospital (Southeast University), Mengchao Hepatobiliary Hospital (Fujian Medical University), Sun Yat-Sen Memorial Hospital (Sun Yat-Sen University), and Eastern Hepatobiliary Surgery Hospital (EHBH). These patients had undergone HCC resection between January 2016 and June 2019. The paired tumor/non-tumor tissue or cell-free DNA (cfDNA) sequencing data of these patients are available. Tumor tissue sequencing data are analyzed to identify MVI-related mutations for the development of a targeted cfDNA sequencing panel. The MVI-related genomic signatures detected in cfDNA sequencing are then used to develop a cfDNA-based MVI-predictive model.

For all enrolled patients, their clinicopathological variables will be utilized to develop the clinical MVI-predictive model.

1. Patient grouping (1) Step 1: Grouping for developing four individual MVI-predictive models.
2. MRI-feature cohort (n=3000): The contrast-enhanced MRI features of these patients will be used to construct both the radiomics and deep learning MVI-prediction models.
3. cfDNA cohort (n=650): The sequencing data of tumor tissue and cfDNA will be used to construct the cfDNA signature-based MVI prediction model.
4. Whole cohort (n=3650): The clinicopathological data of these patients will be used to construct the clinical MVI prediction model.

The final multimodal fusion model will be developed by integrating the predictions from the four individual models.

(2) Step 2: Grouping (re-staging) of early-stage HCC into MVI high-risk and low-risk groups (substages) based on the multimodal MVI prediction.

This analysis will be conducted within the cfDNA cohort, as patients in this group possess all data types required for multimodal MVI prediction, including clinicopathological variables, contrast-enhanced MRI findings, and cfDNA sequencing data.

Based on the risk stratification determined by the multimodal MVI prediction model, patients with early-stage HCC will be re-staged into two substages: the MVI high-risk and MVI low-risk groups. The 5-year OS, 5-year RFS, and recurrence patterns will be compared between the two groups to evaluate the prognostic accuracy of this re-staging system.

(3) Step 3: Grouping for validating the accuracy of re-staging in guiding clinical decisions.

Following the analysis in Step 2, we will evaluate the clinical utility of the re-staging system for surgical decision-making, using the selection of optimal resection margin width as an illustrative example. The MVI high-risk group will be divided into two subgroups: wide-margin resection and narrow-margin resection. The 5-year OS, 5-year RFS, and recurrence patterns will be compared between the two subgroups. The same analysis will be conducted in the MVI low-risk group.

The results will indicate whether the re-staging system can accurately guide the individualized selection of resection margin width in early-stage HCC.

1. Clinical management and follow-up
2. Surgical treatment: Perioperative management and surgical resection are conducted following the clinical guidelines and the standard protocols of each participating center. Specifically, in addition to conventional preoperative assessment, volumetric assessment for future liver remnant (FLR) is performed using three-dimensional reconstruction of imaging studies. All resections are performed with an intention of complete tumor removal with either anatomic or non-anatomic resections. In this study, patients who have undergone a narrow margin resection are reassessed postoperatively to determine whether they are also suitable for a wide margin resection. Three senior hepatic surgeons who are blinded to the prognostic information review the preoperative data of these patients, primarily based on imaging features including tumor size, tumor number, tumor capsule status, intrahepatic location, estimated volume of FLR, degree of cirrhosis, liver function, general performance, as well as their own surgical experience.
3. Poatoperative histopathlogy: Surgicall specimens are routinely examined histopathologically. To ensure the quality of tissue samples in detecting MVI, a seven-site sampling protocol recommended by the Pathology Branch of the Chinese Medical Society (CMS) is used. Both tumor and non-tumor tissues sections are examined to observe MVI. The presence of MVI is identified if there is a microvascular cancer embolus or cancer cell nest in small branches of portal vein or hepatic vein in the adjacent liver tissues, or in large capsular vessels lined by endothelium that is visible only on microscopy. The resection margin width is defined as the nearest distance between the raw surface after resection and the tumor capsule. All studies are independently carried out by three pathologists who arrive at a consensus by discussion if there is any controversy.
4. Follow-up: Patients are followed up regularly after surgery according to a standardized schedule: every 2-3 months for the first 2 years, and every 3-6 months thereafter until each surviving patient had been followed for a least 5 years. During each visit, α-fetoprotein (AFP) and vitamin K absence or antagonist II (PIVKA-II) levels, liver function tests, and abdominal ultrasound are done. Contrast-enhanced CT or MRI of the abdomen is performed every 4 to 6 months, or earlier if clinically indicated. Recurrence/metastasis is defined as appearance of new lesion(s) confirmed on at least two imaging techniques, with or without elevation of tumor markers. Recurrence patterns are categorized by timing (early recurrence: within 2 years post-surgery; late recurrence: beyond 2 years), location (intrahepatic, extrahepatic, or concurrent), and morphological features, including number, size, and site. Local recurrence is defined as recurrence within 2 cm of the resection margin, and distant segment recurrence as recurrence outside the adjacent hepatic segments.
5. Development of clinical MVI-predictive model Preoperative clinical variables include age, sex, total bilirubin (TBIL), alanine aminotransferase (ALT), aspartate aminotransferase (AST), albumin, platelets, prothrombin time (PT), AFP, PIVKA-II, hepatitis B and C serology, HBV-DNA levels, and imaging features such as tumor number, tumor diameter and cirrhosis on preoperative contrast-enhanced MRI.

The clinical model is developed using clinicopathological variables selected based on univariate logistic regression analysis, retaining only those with a P value <0.05. These statistically significant variables are then used as input features to train an XGBoost model. This model is built by iteratively adding decision trees, each designed to correct the errors of previous trees, with hyperparameters such as learning rate, tree depth, and subsample ratio optimized via cross-validation. The final model outputs a continuous risk score (or a calibrated probability) for predicting MVI, capturing non-linear relationships and complex interactions among clinical variables that cannot be readily modeled by logistic regression. Alternatively, feature selection can be refined using XGBoost's built-in gain-based importance scores after an initial univariate filter.

1. Development of MRI-based radiomics and deep learning MVI-predictive models
2. Radiomics model: standardized preprocessing was applied to the multiparametric multisequence MRI images before radiomic feature extraction. Quantitative radiomic features were extracted from the segmented tumor volumes across the available MRI sequences. Least absolute shrinkage and selection operator (LASSO) regression was then performed within each training fold to reduce feature dimensionality and select informative features. The retained features were subsequently entered into a lightweight Transformer-based classifier. Specifically, each selected radiomic feature was projected into a normalized low-dimensional embedding space, and the resulting embeddings were organized as a feature-token sequence. Two stacked Transformer encoder blocks, each comprising dual-head self-attention, were used to capture global dependencies and nonlinear interactions among radiomic features. The learned token-level representations were aggregated using global average pooling and passed through a linear classification layer to generate the predicted probability of MVI..
3. Deep learning model: a customized three-dimensional ResNet-18 architecture was used to learn volumetric representations directly from multisequence MRI data. Each MRI sequence was processed by an independent modality-specific branch comprising an initial three-dimensional convolutional stem, a pooling layer, and four successive residual stages with skip connections. These branches generated modality-specific three-dimensional feature maps at multiple levels of abstraction. The dimension-matched feature maps from different MRI sequences were fused through element-wise addition, followed by normalization and nonlinear activation. Global average pooling was then applied to the fused representation, and the resulting feature vector was passed through a fully connected classification layer to generate the final MVI probability..
4. Development of cfDNA genomic signature-based MVI-predictive model
5. Generation of a tissue-based gene panel for cfDNA targeted sequencing: We first identify MVI-related genomic alterations through Whole-exome sequencing (WES) or targeted sequencing of HCC tumor tissues. Fresh tumor tissues are collected and stored at -80°C, and genomic DNA is extracted using the DNeasy Tissue Kit (Qiagen) followed by fragmentation to 200-400 bp. WES libraries are prepared from 1 μg of DNA with the SureSelect XT system (Agilent) and sequenced on an Illumina HiSeq 2000 platform (2×100 bp reads). In parallel, targeted sequencing is performed on the same platform using a commercial 123-gene HCC panel (Baodeng Bio), achieving a minimum coverage depth of 1000×. Bioinformatic analysis is performed by aligning all sequencing data to the hg19 reference genome. Somatic mutations (SNPs and indels) are called from WES and targeted sequencing data using MuTect2. Germline mutations are identified from WES data via MutSigCV with a false discovery rate (FDR) <0.001 and then subjected to GO and KEGG pathway enrichment analysis. Genes within significantly enriched pathways (P<0.05) are retained as candidates. To further screen for MVI-related genes, the maximum mutant allele frequency (maxAF) of each somatically mutated gene is calculated and evaluated using logistic regression, with genes showing P<0.05 in targeted sequencing data and P<0.1 in WES data considered relevant. The resulting candidate genes from both somatic and germline analyses are consolidated to design a focused cfDNA sequencing panel.
6. Development of a cfDNA genomic signature-based MVI-predictive model: Peripheral blood (10 mL per patient) is collected pre-operatively and processed within 1 hour. Plasma is separated by centrifugation (1,600g for 10 min) and further cleared by high-speed centrifugation (20,000g for 10 min) to remove cellular debris. cfDNA is extracted from 3-5 mL plasma using the QIAamp Circulating Nucleic Acid Kit (Qiagen) and quantified with a Qubit fluorometer. White blood cells (WBCs) are also collected from the same blood samples; their genomic DNA is extracted with the DNeasy Blood Kit (Qiagen), fragmented, and used for library construction with a KAPA kit. WBC libraries are captured using the Accu-Act panel (AccuraGen) and sequenced on an Illumina HiSeq 2500 system, achieving a minimum coverage of 500×. cfDNA libraries are prepared and sequenced on the same platform using the Firefly system (AccuraGen) with the MVI-PG37 panel, yielding an average coverage of approximately 7000×.

To build the predictive model, background noise introduced by random NGS errors is first removed from cfDNA and WBC data using an AccuraGen proprietary algorithm. Somatic mutations in cfDNA are identified by cross-referencing tumor-tissue mutations with germline variants from WBCs; the maxAF of each somatically mutated gene is used as a quantitative somatic signature. Germline SNPs and indels detected in WBCs are recorded as binary (presence/absence) germline signatures. Significant signatures are selected through univariable logistic regression followed by forward stepwise multivariable selection, optimized for maximum area under the receiver operating characteristic curve (AUC). These selected somatic and germline signatures are finally integrated into a cfDNA-based model to predict MVI status. Candidate algorithms, including XGBoost, penalized logistic regression and gradient-boosted decision trees (GBDT), will be evaluated within the training data using nested cross-validation, and the final algorithm will be selected according to prespecified performance.

1. Development of multimodal MVI-predictive model As noted earlier, four individual models are developed to preoperatively assess MVI risk: a clinical model, an MRI-based radiomics model, an MRI-based deep learning image model, and a cfDNA mutation model. To enable standardized integration across heterogeneous data modalities, raw outputs from the four individual models are calibrated into risk probabilities ranging from 0.0 to 1.0 using Platt scaling or isotonic regression. Platt scaling fits a logistic regression model to the raw scores, assuming a sigmoidal relationship, and is particularly effective when the uncalibrated outputs are from support vector machines or similarly distributed models. In contrast, isotonic regression learns a nonparametric, piecewise-constant monotonic function from the data, making no prior assumption about the shape of the calibration curve but requiring a larger calibration set to avoid overfitting. This calibration step ensures that scores from different sources are comparable and well-suited for downstream combination, especially when model outputs are not naturally probabilistic. The calibrated probabilities will be clipped to avoid boundary values of 0 and 1 and transformed to logits before being entered into an L2-regularized logistic stacking model. The fusion coefficients of the final multimodal system will be estimated exclusively from out-of-fold predictions. After all preprocessing, calibration, fusion and threshold-selection procedures have been locked, the complete pipeline will be refitted using the full training cohort and evaluated once in the temporally independent validation cohort, which will remain inaccessible throughout model development. .
2. Development of re-staging system for early-stage HCC The total risk probability of MVI prediction of each patient is calculated, and ROC analysis is used to calculate the optimal cutoff value of the multimodal model to distinguish between high and low-risk for MVI by maximizing the Youden index (i.e., sensitivity+specificity-1). Using this cutoff value, patients with early-stage HCC with a predicted MVI probability at or above the cutoff determined by the maximum Youden index will be classified as high risk, whereas those below the cutoff will be classified as low risk.
3. Study endpoints The primary outcome is investigator-assessed RFS, defined as the time from surgery to the first diagnosis of recurrence or death from any cause. Event-free patients are censored at the last follow-up. Secondary outcomes included OS and recurrence patterns. OS is measured from surgery to death from any cause, with surviving patients censored at the last follow-up. Recurrence patterns are categorized by timing (early recurrence, ≤2 years after surgery; late recurrence, >2 years), location (intrahepatic, extrahepatic, or concurrent), and morphological features, including number, size, and site. Local recurrence was defined as recurrence within 2 cm of the resection margin, and distant segment recurrence as recurrence outside the adjacent hepatic segments.
4. Statistical analysis Continuous variables are presented as median (interquartile range [IQR]) and compared using the Kruskal-Wallis tests for comparisons among three groups or the Mann-Whitney U test between two groups. Categorical variables are presented as number (%) and compared usingχ2 test or Fisher's exact test. A multiple imputation approach is planned to handle possible missing data. The performance of the developed models is evaluated using the area under the ROC curve (AUC), sensitivity, specificity, positive predictive value, negative predictive value, concordance index (C-index), and calibration analysis. Interaction terms between variables that can predict MVI are analyzed using the logistics regression. The net benefits that evaluated the potential clinical utility of cfDNA signatures or MVI-prediction models are assessed by the decision curve analysis (DCA). The Delong test is used to compare differences between two ROC curves. RFS and OS are estimated using the Kaplan-Meier method and log-rank test. The univariable and multivariable Cox proportional hazards model is used to estimate the hazard ratios (HRs) and 95% confidence intervals (CIs) to identify independent risk factors of RFS and OS. Variables with a p value of <0.05 in the univariable COX analysis are selected into the multivariable analysis, and variables with a p value of <0.05 in the multivariable COX analysis are retained in the final model. Subgroup analyses for model performances are performed based on the factors, including tumor size, hepatitis B surface antigen (HBsAg) status, and cirrhosis. A P value of <0.05 is considered statistically significant, unless otherwise specified. Statistical analysis is performed using the python programming language version 2.7 (https://www.python.org/) and the R software version 3.1.1 (http://www.r-project.org/).
5. Reporting standards Our results will be reported per the Strengthening The Reporting of Observational Studies in Epidemiology (STROBE) and Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines.