The recurrence problem Despite curative hepatectomy, HCC recurs in 50-70% of patients within 5 years. Microvascular invasion (MVI) - cancer cells invading small blood vessels near the tumor - is one of the strongest predictors of recurrence, present in 15-57% of surgical specimens even for small HCC.
MVI is only diagnosed after surgery MVI is defined by microscopic tumor cell invasion of portal vein, hepatic vein, or capsular vessels and can only be definitively detected on pathologic examination of the resected specimen. This means surgeons cannot currently use MVI status to guide the extent of resection, transplant eligibility, or neoadjuvant treatment decisions.
Radiomics as a preoperative tool Radiomics extracts hundreds of quantitative imaging features from MRI or CT that capture tumor heterogeneity, texture, and shape beyond what radiologists can visually assess. This study applies multi-scale (tumor, peritumoral, liver parenchyma) and multi-parametric (DWI, arterial, portal venous, hepatobiliary phase) radiomics to predict MVI before surgery.
Study goal This study developed and validated radiomics-based nomograms to predict MVI status and recurrence-free survival (RFS) preoperatively in a cohort of 356 patients with solitary HCC less than or equal to 5 cm who underwent gadoxetate disodium (Gd-EOB-DTPA)-enhanced MRI.
Patient cohort 356 patients with pathologically confirmed solitary HCC less than or equal to 5 cm who underwent preoperative Gd-EOB-DTPA (gadoxetate disodium) MRI were retrospectively enrolled. Patients were split into training (n=250) and validation (n=106) cohorts. MVI was graded as M0 (no invasion), M1 (up to 5 vessels within 1 cm), or M2 (greater than 5 vessels or beyond 1 cm).
Multi-parametric MRI protocol The study used all major Gd-EOB-DTPA MRI sequences: diffusion-weighted imaging (DWI), arterial phase (AP), portal venous phase (PVP), transitional phase (TP), and hepatobiliary phase (HBP). This is more comprehensive than most prior radiomics studies that focused only on hepatobiliary phase images.
Multi-scale regions of interest Radiomics features were extracted from three spatial scales: the entire tumor volume (intratumoral), the peritumoral area within 10 mm of the tumor margin (an area rich in invasive cells often neglected in prior studies), and randomly selected liver parenchyma regions. This multi-scale approach captures tumor-environment interaction.
Statistical modeling Two analytical approaches were compared: random forest (RF) and logistic regression (LR). LASSO (Least Absolute Shrinkage and Selection Operator) regularization was used for feature selection. Final nomograms combined radiomics scores with clinical and radiologic predictors identified by multivariate analysis.
Independent predictors of MVI Multivariate analysis identified elevated AFP (greater than 400 ng/mL), elevated total bilirubin, high radiomics score, peritumoral enhancement on arterial phase, and incomplete or absent capsule enhancement as independent risk factors for MVI. The radiomics score was the predominant independent predictor.
Nomogram AUC performance The MVI nomogram using random forest achieved an AUC of 0.920 (95% CI: 0.861-0.979) in the validation cohort - higher than any previously reported preoperative MVI prediction model. The logistic regression-based nomogram achieved AUC 0.879 (95% CI: 0.820-0.938), also strong performance.
Peritumoral features matter Inclusion of peritumoral radiomics (from the region within 10 mm of the tumor) significantly improved prediction performance compared to intratumoral features alone. This highlights that the microenvironment immediately surrounding the tumor contains critical invasion information not captured by tumor-only analysis.
Multi-phase advantage Features from both DWI and dynamic contrast-enhanced sequences (arterial, portal venous, hepatobiliary phases) contributed complementary information. No single phase was sufficient - the combination of multiple imaging parameters reflecting different aspects of tumor biology was key to high accuracy.
MVI strongly predicts recurrence Pathologic MVI was the primary independent risk factor for postoperative recurrence. The 5-year recurrence-free survival (RFS) rate was 68.4% overall. MVI-positive patients (M2 and M1) had median RFS of 30.5 months (11.9 months for M2, 40.9 months for M1), versus greater than 96.9 months for MVI-negative (M0) patients (p less than 0.001).
The RFS nomogram An independent RFS prediction nomogram was developed incorporating age, histologic MVI status, alkaline phosphatase, and alanine aminotransferase, achieving AUC of 0.654 in the validation cohort. While modest, this nomogram provides preoperative RFS guidance.
Preoperative MVI substitutes for histologic MVI The key translational finding: replacing histologic MVI with preoperatively predicted MVI (MVI-RF) from the radiomics nomogram achieved comparable accuracy in MVI stratification and comparable RFS prediction. This validates the radiomics nomogram as a practical substitute for post-surgical pathological MVI assessment.
Clinical upstaging implications Identifying MVI-positive patients preoperatively allows surgeons to pursue wider resection margins (shown to improve outcomes in MVI-positive HCC), consider neoadjuvant therapies, or prioritize patients for closer post-surgical surveillance.
Preoperative surgical planning Knowing MVI status before surgery enables surgeons to plan appropriately wide resection margins for MVI-positive patients, which is associated with reduced recurrence rates. This is currently impossible without the radiomics tool, since MVI is only known after the specimen is examined.
Transplant prioritization MVI-positive patients may be better candidates for liver transplantation over resection given the higher recurrence risk with surgery. Preoperative identification of MVI could shift treatment recommendations toward transplantation when feasible.
Adjuvant therapy decisions High MVI burden (M2 grade) patients have extremely poor outcomes (median RFS 11.9 months). Identifying these patients before surgery could prompt enrollment in adjuvant trials or earlier initiation of adjuvant systemic therapy to address micrometastatic disease.
Non-invasive monitoring Serial preoperative MRI radiomics assessment could potentially track changes in MVI risk during bridging or downstaging therapies (TACE) before transplantation, providing dynamic rather than static risk stratification.
Retrospective single-center design This is a retrospective study from a single Chinese institution, which may limit generalizability. External validation in diverse populations including non-HBV-predominant HCC (e.g., alcohol-related or NAFLD-related) is needed before broad clinical adoption.
Manual segmentation burden Radiomics feature extraction requires manual or semi-manual tumor segmentation by trained radiologists. The time investment and inter-observer variability in segmentation remain practical barriers to clinical implementation.
MRI protocol standardization The study used a specific Gd-EOB-DTPA protocol at a 1.5T scanner. Validation across different scanner types, field strengths, and contrast agents is required, as radiomics features are sensitive to image acquisition parameters.
Model interpretability Random forest models, while achieving the highest AUC, are less interpretable than logistic regression nomograms. Clinical adoption may favor logistic regression-based nomograms for their transparency, despite slightly lower performance.
Automated segmentation Development of automated deep learning-based tumor segmentation algorithms will eliminate the manual segmentation bottleneck and enable radiomics analysis as part of routine clinical workflow, making the approach scalable.
Prospective validation Prospective multi-center studies with standardized MRI protocols and centralized radiomics analysis pipelines are needed to validate the nomograms across diverse patient populations and scanner platforms before regulatory approval.
Integration with molecular biomarkers Combining radiomics with liquid biopsy biomarkers (ctDNA, AFP variants) or genomic tumor profiling could create multi-modal prediction tools with higher accuracy than imaging or molecular markers alone.
Extending beyond small HCC The current study focused on solitary HCC less than or equal to 5 cm. Validating similar radiomics approaches for larger tumors, multifocal HCC, and other liver malignancies (intrahepatic cholangiocarcinoma) could broaden clinical impact.