Hepatocellular carcinoma (HCC) is a major global health challenge with a rising incidence worldwide (1). HCC development is closely related to chronic liver disease with viral hepatitis, alcoholic and non-alcoholic steatohepatitis as leading etiologies (1). Due to the reliance on tumor burden and the functional hepatic reserve for determining patient treatment and outcomes, managing HCC proves exceptionally challenging (2). The complexity of this heterogenous disease and its treatment is reflected in markedly variable outcomes following potentially curative therapy such as surgical resection, liver transplantation, or local ablation (3). Patients undergoing these treatments typically exhibit a 5-year survival rate of approximately 62-70% (4, 5). Moreover, HCC recurrence in patients after potentially curative treatment remains a major burden with rates up to 70% within 5 years after treatment (6). Microvascular invasion (MVI), characterized by the microscopic presence of tumor cells in hepatic vessels (arteries, hepatic vein, and portal vein) lined with endothelial cells, stands as the most crucial determinant of disease recurrence and long-term survival (7). Unfortunately, MVI can only be histopathologically diagnosed based on the resected surgical specimen and therefore its use to guide personalized treatment strategies remains limited. Recently, the prediction of MVI before surgery has gained increasing attention, with several promising noninvasive methods utilizing imaging features or tumor markers embedded into risk models (8–12). However, current evidence is primarily characterized by complex risk models involving multimodal biomarkers, or restriction of imaging modalities to either magnetic resonance imaging (MRI) or computed tomography (CT) (13, 14). Furthermore, the vast majority of risk models were developed in selective subsets of HCC patients (i.e., predominantly viral hepatitis) with significant imbalances of the number of predictor variables and high MVI rates in the cohorts hampering its transferability to the clinical routine (12, 15, 16).

In this study, we aimed to develop a noninvasive digital biopsy risk model to predict MVI using preoperative imaging features and assess its prognostic outcome in patients undergoing hepatectomy for HCC.

A total of 217 patients underwent hepatectomy for hepatocellular carcinoma at both institutions during the study period ( Figure 1 ). Of these, 37 patients (17%) had a histopathological positive MVI. The baseline characteristics of patients with and without MVI are detailed in Table 1 . In the MVI-positive group, more patients showed lesions exceeding 5 cm (68% vs. 43%, P = 0.030) and required major hepatectomies (46% vs. 17%, P = 0.030) as compared to patients in the MVI-negative group. Other characteristics were well-balanced between the groups.

In the present dual-center study, we developed a straightforward scoring system to predict MVI in patients with HCC using distinct preoperative radiologic features. Our digital biopsy model included the presence of a non-smooth peritumoral margin, intratumoral arterial enhancement, and irregular peritumoral hyperenhancement. The presence of at least two out of three radiologic features was associated with high sensitivity (87%) and specificity (88%) for predicting MVI before hepatectomy. Moreover, the survival outcomes stratified by our digital biopsy risk model achieved a significant difference of overall and recurrence-free survival in patients with predicted high risk versus low risk of MVI.

MVI is defined as the microscopic presence of cancer cells in hepatic vessels, indicating aggressive tumor biology (7). On preoperative imaging, a non-smooth peritumoral margin implies an upfront sign of tumor aggressiveness characterized by tumor protrusion into peritumoral areas (26). A meta-analysis on the predictive value of a nonsmooth peritumoral margin revealed that a non-smooth peritumoral margin is associated with a diagnostic odds ratio of >20 for MVI (27). Three other meta-analyses on the impact of different preoperative imaging features to predict MVI determined that a nonsmooth peritumoral margin is an important independent predictor of MVI (28–30). In the present study, we confirmed that this marker was the strongest predictor of MVI.

The presence of internal arteries in the portal phase is another substantial radiologic marker of MVI. Previous studies reported that internal arteries in HCC are correlated with angiogenesis and cellular proliferation which in turn results in tumor progression (31, 32). In 2007, a radiogenomic biomarker to predict MVI was developed in 28 patients with HCC and further validated in a cohort of 157 patients (29% with MVI) (20, 33). This radiogenomic biomarker was based on the correlation of two combined radiologic features (i.e., presence of internal arteries and peritumoral hypodense halo) with angiogenesis gene expression patterns and resulted in a sensitivity and specificity of 76% and 94% to predict MVI (20, 33). Some studies confirmed the predictive value of these two combined radiogenomic features (8, 34), while other studies including larger patient cohorts reported conflicting results (35, 36) and discovered an even higher predictive and prognostic value for internal arteries compared to peritumoral hypodense halo (12). Thus, in our study, we evaluated no combined imaging features and depicted a high predictive value of internal arteries for MVI, while peritumoral hypodense halo failed statistical significance to predict MVI.

Another characteristic imaging feature of advanced HCC is the presence of a radiologic tumor capsule (referred to pseudocapsula or peritumoral hyperenchancement), found in 70% of HCC cases (37). While the absence of a pseudocapsula might indicate an early HCC, an irregular or incomplete pseudocapsula is associated with MVI (diagnostic odds ratio of 1.85) according to a meta-analysis (38). In the present study, an irregular pseudocapsula was one of the three independent imaging features associated with MVI.

So far, a plethora of risk models exists in the literature, incorporating imaging features to predict MVI (11, 39, 40). However, the majority of available risk models were developed in patients with viral HCC having a high incidence of MVI. These models included multiple candidate variables in multivariate analyses based on the “rule of thumb”, without considering sample size considerations (8, 9, 12, 16). Renzulli et al. identified three “worrisome” features (i.e., nonsmooth tumor margin, the radiogenomic features (combination of internal arteries and hypodense peritumoral halo), and irregular pseudocapsula) to be associated with MVI in a total of 140 patients (64% with MVI and 6 candidate variables) with a c-index of 0.85 and 0.90 (8). However, these features were not tested in a multivariate analysis (8). Similarly, Min et al. described a diagnostic model in a total of 100 patients (39% with MVI) including four radiologic features (i.e., non-smooth margin, irregular pseudocapsula, peritumoral hyperenhancement, peritumoral hypointensity) with a c-index of 0.80, again without multivariate testing (9). Lee et al. developed a 6-point risk model including two radiologic features (i.e., peritumoral arterial enhancement, peritumoral hypointensity) and two serological biomarkers in a total of 276 patients (28% with MVI and 15 candidate variables) with a c-index of 0.87 (16). Recently, Jiang et al. presented a complex 10-point risk model that outperformed the models of Renzulli et al., Min et al., and Lee et al. in a comparative analysis. However, the risk model by Jiang et al. was developed in 319 patients (47% with MVI), evaluating 22 candidate variables in a multivariate analysis. Hence, the current risk models in literature are at high risk of statistical overfitting owing to the high number of candidate variables and the selection of MVI patients (MVI incidence ranging between 27% - 64%) (24). The major strength of our digital biopsy risk model is that we performed a formal sample size calculation and adjusted our analysis by controlling for several confounders (i.e., etiology of disease, lesion pattern, age, sex) which were not addressed in previous studies (8, 9, 16, 20, 33). Our digital biopsy risk model achieved a high discrimination value (c-index of 0.87) and yielded excellent calibration metrics. Notably, we included only three imaging features on CT or MRI, making it highly applicable in the daily routine compared to other models comprising more features with or without additional serum analyses as well as restrictions on the imaging modality (i.e., MRI or CT) (8–12). Additionally, the prognostic utility of our risk model was further proven by stratified survival analyses. Remarkably, patients with a high risk of MVI on the digital biopsy risk model reflected an even worse prognosis in terms of overall survival compared to patients with histopathologically proven MVI. This finding is in line with other reports in literature, indicating a higher prognostic performance of pretreatment radiologic features as compared to histopathological MVI (41, 42). Therefore, our risk model provides a “digital biopsy” and may represent an additional noninvasive armamentarium to facilitate personalized HCC treatment strategies and improve patient outcome. To this end, patients with high risk of MVI might benefit from neoadjuvant local (i.e., chemoembolization, radioembolization) or systemic treatment prior surgical resection. Given that patients with high risk of MVI recur more frequently, the digital biopsy model could also be helpful to stratify candidates for liver transplantation and, thus, impacting organ allocation policies.

There are some limitations to our study. This is a retrospective prognostic study with a potential selection bias. To increase the generalizability of our results, we performed a rigorous methodology with adjusted analyses, and imaging features were evaluated by local radiologists at each center. Still, our study lacks an external validation cohort, and the findings of our study need to be tested in a separate HCC population.

In conclusion, we developed and internally validated a robust and reliable prediction model of MVI.