Hepatocellular carcinoma (HCC) is a malignancy with third highest world mortality rate (8.3%), after lung cancer (18%) and colorectal cancer (9.4%) (1). The main treatment of HCC is surgical resection, but recurrence is common, with a five-year recurrence rate of up to 40%–70%. The 5-year survival rate is only 18% (2). Microvascular invasion (MVI) is one of the most important prognostic factors for HCC after surgical treatment, and it has been established as a risk factor for early recurrence and poor outcome. The term MVI refers to the cancer cell nests that are established within the lining of blood vessels by microscopic endothelial cells. It is considered as a sign of strong tumor invasion ability, and it could only be diagnosed through pathology at present. MVI-positive patients often require improved prognosis by expanding surgical margins, and patients with microvascular infiltration are also considered unsuitable for liver transplantation (3–5). Therefore, developing a method for non-invasive prediction of microvascular invasion is necessary to guide the treatment of HCC.

The imaging characteristics of HCC, such as a non-smooth tumor margin, arterial peritumoral enhancement, and peritumoral hypo-intensity on hepatobiliary phase imaging (HBP), have been confirmed to be noninvasive imaging biomarkers for MVI prediction (4–7). However, such qualitative studies are vulnerable to subjective factors, image quality, and interobserver variation. Therefore, more objective quantitative methods are needed to predict MVI. In 2012, Lambin et al. (8) proposed the concept of radiomics, where medical images are converted into useful data by using high-throughput quantitative features to predict the disease treatment efficacy and prognosis. Using radiomics to predict MVI in HCC is a major research area in recent years. Many studies with satisfactory results have been conducted (9–11). Xu’s (10) study has achieved predicted satisfactory results by developing a radiomics nomogram model on the basis of computed tomography (CT). MRI examination technology has the advantage of multimodal/multisequence imaging and high soft-tissue resolution. Theoretically MRI multicolumn multimodal imaging provides more characteristic elements. Therefore, in the present study, predictive models based on radiomic features in MRI were developed for predicting MVI.

MicroRNAs (miRNAs) are a kind of endogenous, non-coding RNAs. Thousands of miRNAs play a role in regulating various molecular biological processes by inhibiting the translation of different messenger RNAs (mRNAs) in the cell (12). A dysregulation of miRNAs is often associated with malignancy, and it regulates the proliferation, migration, invasion, and development of tumors in HCC by promoting or suppressing them (13, 14). Previous studies have shown that combining radiomics and genomics could remarkably improve the performance of predictive models (15, 16). Zhou (17) et al. screened 7 plasma miRNAs (miRNAs) out of 723 HCC-associated miRNAs (miR-122, miR-192, miR-21, miR-223, miR-26a, miR-27a, miR-801), which had high diagnostic performance in the early diagnosis of hepatocellular carcinoma.Therefore, we extracted these 7 mi-RNAs from the patients’ plasma, but among them, miR-192 and miR-801 had large differences in expression, and the data were not stable enough to be screened out, so only 5 mi-RNAs (miR-122, miR-21, miR-223, miR-26a, miR-27a) were included in the analysis. Therefore, we measured these miRNAs and explored their relationship with HCC microvascular invasion, and combined them with radiomics to explore the performance of the joint model. Moreover, routine laboratory tests for HCC and radiological characteristics based on MRI were added; the independent risk factors for MVI were determined through multivariate logistic regression, combined radiomics, genomics, and clinico-radiological factors; predictive models were established; and the performance of these models was verified.

This study aimed to develop and validate a radiogenomics nomogram model for preoperative prediction of MVI in HCC. The nomogram is helpful for clinicians to assist in determining individual therapeutic strategies for patients with HCC.

In this study, we have successfully developed and rigorously validated a multi-omics nomogram prediction model, which integrates MRI-derived radiomics, radiological features, and miRNA-based genomics. The resulting radiogenomic nomogram has demonstrated excellent performance in accurately predicting microvascular invasion in HCC, thus providing a non-invasive yet reliable clinical method for preoperative prediction.

The expression of miR-21, miR-27a, and miR-122 in the MVI⁺ group was upregulated compared with that in the MVI⁻ group, whereas the expression of miR-26a and miR-223 was downregulated. These differences were all statistically significant. By contrast, the differences in conventional laboratory indicators, such as AFP, were not statistically significant between the MVI⁺ group and the MVI⁻ group, indicating that the miRNAs extracted in the plasma of patients with HCC were more meaningful in suggesting MVI than the conventional laboratory indicators. Further univariate and multivariate analyses showed that miR-21 had better correlations than other clinical laboratory tests. Studies have shown that miR-21 is one of the most expressed miRNAs in liver diseases, such as nonalcoholic fatty liver disease (21). Ladeiro et al. (22) found that miR-21 was significantly overexpressed in HCC, as compared to benign tumor or non-neoplastic liver tissue. The maladjusted expression of miRNA could be used as a biomarker, and it could be detected in the plasma of patients. Studies have shown that miR-21 plays a role in promoting HCC growth invasion, distant metastasis, and other links (21). Furthermore, this study confirmed that among other miRNAs detected in plasma, only miR-21 emerged as an independent risk factor for MVI in HCC. The miR-21 model surpassed the rad score and maximum tumor length models in predicting MVI, and the differences were statistically significant in the validation cohort (P<0.05). These findings indicate that the miR-21 model performs well in preoperative prediction of MVI in HCC. Conventional clinical laboratory indicators, including tumor markers such as AFP, were all excluded in the univariate analysis due to their relatively poor correlation compared to other variables.

On the basis of the morphological characteristics of MRI, previous meta-analysis studies have found that some of MRI signs were significantly associated with MVI, including larger tumors (> 5 cm), rim arterial enhancement, arterial peritumoral enhancement, non-smooth tumor margin, and multiple lesions (7). The present study incorporated these MR morphological features into the analysis and found that only the maximum tumor length is an independent predictor of MVI in HCC. Increased tumor volume led to increased contact between the tumor and adjacent normal liver tissues, thereby promoting the formation of microvessels.

A total of 7045 features were extracted from five different MRI sequences, and then 11 radiomic features were screened out by SelectKBest and LASSO. Interestingly, these features were all derived from DWI and DP sequences. A previous study (23) found that primary radiomic signatures extracted from delayed-phase sequences were associated with MVI. Zhang et al. (24) compared the performance of different MRI sequences to predict MVI and found the performance of DP to be the best (AUC = 0.806). The present study also found the value of DP to be the best in predicting MVI, which could be explained by the fact that tumors continue to release a large number of angiogenic factors that promote tumor angiogenesis and change tumor perfusion, resulting in the differences between MVI⁻and MVI⁺ being more easily shown in DP sequences (25). DWI also has a satisfactory performance, as confirmed by some previous studies (26, 27). This finding could be explained that the minimum value of the apparent diffusion coefficient of DWI could reflect the densest tumor, the most abundant neovascularization, and the most active tumor proliferation. The hepatobiliary phase of specific contrast agents has been reported to make an important contribution to suggesting MVI (7). It is important to note that when hepatocyte-specific agents are used, DWI is usually scanned after contrast, and the ability and contribution of DWI may differ, which needs to be proven by further research.

After rigorous data analysis and model training, we have successfully developed a nomogram model that comprehensively incorporates various risk factors. In stringent tests using training and validation datasets, the model exhibited outstanding predictive performance, with AUC values reaching 0.942 and 0.900, significantly surpassing other single-factor models. The nomogram model demonstrated satisfactory performance in predicting microvascular invasion in HCC. Looking back at previous studies, although nomogram models based on radiomics and clinical factors have shown promising predictive capabilities, most of them still have limitations in performance, with AUC values generally ranging from 0.801 to 0.861 (28–30). Notably, these models often involve numerous risk factors, whereas our nomogram model incorporates only three independent risk factors, highlighting the advantages of multi-omics approaches in data processing and model development.

In previous explorations, such as the study by Banerjee et al. (28), they delved into the radiogenomics of MVI in liver cancer, innovatively developing a novel imaging biomarker called radiogenomic venous invasion (RVI) by combining venous invasion genes in hepatocellular carcinoma with dynamic contrast-enhanced CT. This achievement has achieved significant results in predicting MVI and prognosis. Similarly, Taouli et al. (29) also conducted in-depth research on the imaging characteristics and genomic data of hepatocellular carcinoma, successfully identifying imaging features related to aggressive hepatocellular carcinoma genes through a combination of preoperative CT or MR examinations and transcriptomic analysis.

However, our study adopted a more unique and precise approach. We directly used miRNAs closely related to liver cancer as variables in logistic regression analysis, screening out miRNAs species independently associated with MVI in liver cancer through rigorous univariate and multivariate analysis. Subsequently, we combined these crucial miRNAs with radiomics and clinical radiological features to construct an efficient and accurate prediction model, achieving satisfactory preoperative prediction of MVI. This achievement provides a powerful tool or method for the accurate prediction of microvascular invasion in hepatocellular carcinoma, guiding clinical decision-making, optimizing treatment plans, and ultimately improving the survival rate and quality of life of patients.

The radiogenomic nomogram exhibited promising preoperative predictive capabilities and clinical decision-making implications in forecasting microvascular invasion (MVI) in hepatocellular carcinoma (HCC). This model holds the potential to emerge as a biomarker for MVI in HCC in the future, though its efficacy necessitates further validation through extensive studies encompassing larger sample sizes from multiple centers.

The data analyzed in this study is subject to the following licenses/restrictions: Not publicly available due to privacy of laboratory data. Requests to access these datasets should be directed to QZ, wfcyzhangqian@126.com.

The studies involving humans were approved by Institutional Review Board of Yantai Yuhuangding Hospital, Affiliated Hospital of Qingdao University. The studies were conducted in accordance with the local legislation and institutional requirements. The human samples used in this study were acquired from primarily isolated as part of your previous study for which ethical approval was obtained. Written informed consent for participation was not required from the participants or the participants’ legal guardians/next of kin in accordance with the national legislation and institutional requirements.

GH: Formal Analysis, Investigation, Methodology, Writing – original draft. JQ: Data curation, Writing – review & editing. JG: Data curation, Writing – review & editing. YC: Data curation, Visualization, Writing – review & editing. FW: Data curation, Formal Analysis, Writing – review & editing. HZ: Writing – review & editing. HZ: Conceptualization, Data curation, Methodology, Writing – review & editing. XW: Data curation, Supervision, Writing – review & editing. HM: Investigation, Writing – review & editing. HX: Formal Analysis, Project administration, Writing – review & editing. CX: Supervision, Writing – review & editing. NL: Writing – review & editing, Software, Supervision. QZ: Funding acquisition, Supervision, Writing – review & editing.