Evaluating microvascular invasion in hepatitis B virus-related hepatocellular carcinoma based on contrast-enhanced computed tomography radiomics and clinicoradiological factors

Abstract

BACKGROUND

Microvascular invasion (MVI) is a significant indicator of the aggressive behavior of hepatocellular carcinoma (HCC). Expanding the surgical resection margin and performing anatomical liver resection may improve outcomes in patients with MVI. However, no reliable preoperative method currently exists to predict MVI status or to identify patients at high-risk group (M2).

AIM

To develop and validate models based on contrast-enhanced computed tomography (CECT) radiomics and clinicoradiological factors to predict MVI and identify M2 among patients with hepatitis B virus-related HCC (HBV-HCC). The ultimate goal of the study was to guide surgical decision-making.

METHODS

A total of 270 patients who underwent surgical resection were retrospectively analyzed. The cohort was divided into a training dataset (189 patients) and a validation dataset (81) with a 7:3 ratio. Radiomics features were selected using intra-class correlation coefficient analysis, Pearson or Spearman’s correlation analysis, and the least absolute shrinkage and selection operator algorithm, leading to the construction of radscores from CECT images. Univariate and multivariate analyses identified significant clinicoradiological factors and radscores associated with MVI and M2, which were subsequently incorporated into predictive models. The models’ performance was evaluated using calibration, discrimination, and clinical utility analysis.

RESULTS

Independent risk factors for MVI included non-smooth tumor margins, absence of a peritumoral hypointensity ring, and a high radscore based on delayed-phase CECT images. The MVI prediction model incorporating these factors achieved an area under the curve (AUC) of 0.841 in the training dataset and 0.768 in the validation dataset. The M2 prediction model, which integrated the radscore from the 5 mm peritumoral area in the CECT arterial phase, α-fetoprotein level, enhancing capsule, and aspartate aminotransferase level achieved an AUC of 0.865 in the training dataset and 0.798 in the validation dataset. Calibration and decision curve analyses confirmed the models’ good fit and clinical utility.

CONCLUSION

Multivariable models were constructed by combining clinicoradiological risk factors and radscores to preoperatively predict MVI and identify M2 among patients with HBV-HCC. Further studies are needed to evaluate the practical application of these models in clinical settings.

Key Words: Radiomics; Contrast-enhanced computed tomography; Hepatocellular carcinoma; Microvascular invasion; Hepatitis B virus

Core Tip: Preoperative microvascular invasion (MVI) prediction in patients with hepatocellular carcinoma (HCC) is paramount for guiding surgical decisions. Based on contrast-enhanced computed tomography radiomics and clinicoradiological factors, our models offer valuable predictive capabilities for MVI and high-risk groups among those with hepatitis B virus-related HCC, supporting personalized surgical planning and clinical management.

INTRODUCTION

Primary liver cancer is a highly malignant tumor, with hepatocellular carcinoma (HCC) comprising 75%-85% of cases[1]. Hepatitis B virus (HBV) infection is the primary risk factor for HCC in China[2]. Curative treatment options for early-stage HCC include surgical resection, liver transplantation, and ablation therapy, while surgical resection is the most essential approach for achieving long-term survival. However, the 5-year recurrence rate following liver resection remains alarmingly high, reaching up to 70%[3].

Microvascular invasion (MVI) is an independent risk factor for postoperative recurrence, with an incidence ranging from 15% to 57.1% in patients with HCC[4-7]. One study revealed that patients with HCC and MVI who underwent liver resection with a wide surgical margin (≥ 1 cm) experienced significantly lower recurrence rates and higher survival rates than those with a narrow margin (< 1 cm)[8]. Moreover, our center’s research demonstrated that anatomical liver resection prolonged postoperative survival only among patients in the high-risk group (M2) compared to non-anatomical liver resection, consistent with results from Zhang et al[9] and Zhao et al[10]. Consequently, surgical strategies such as wide resection margins and anatomical liver resection can enhance outcomes in patients with HCC and MVI. Therefore, accurate preoperative prediction of MVI and identification of those at high risk (M2) are crucial.

Numerous recent studies have focused on the preoperative estimation of MVI. Beaufrère et al[11] identified a 6-gene signature strongly associated with MVI, which holds promise for clinical application through routine tumor biopsy and could be integrated into therapeutic strategies. However, due to the known limitations of HCC biopsy (including its invasiveness and tumor seeding), this method for detecting MVI has not gained widespread recognition.

Relevant research shows that preoperative imaging and laboratory examinations can predict MVI. Lei et al[12] developed a nomogram integrating seven clinicoradiological factors, demonstrating the excellent preoperative predictive ability for MVI patients with HBV-HCC within the Milan criteria. Lee et al[13] investigated the predictive value of various imaging signs from gadoxetic acid-enhanced magnetic resonance imaging (MRI) for MVI. Their results showed that arterial peritumoral enhancement, non-smooth tumor margins, and peritumoral hypointensity were independently associated with MVI.

Radiomics, a noninvasive approach to extracting quantitative features from medical images, has become increasingly valuable in modern medicine[14]. It provides detailed information that enhances diagnostic accuracy. Recently, radiomics models based on computed tomography (CT), MRI, etc., have been reported to accurately predict MVI. Xia et al[15] developed a radiomics model to predict MVI status based on preoperative multiphase CT images. The model performed well, with an area under the curve (AUC) of 0.86 in the internal and 0.84 in the external test set. Yang et al[16] constructed a nomogram based on clinicoradiological factors and radiomics features from hepatobiliary phase images and demonstrated satisfactory preoperative predictive ability for MVI. The study by Li et al[17] demonstrated that the texture features in positron emission tomography-CT were significantly associated with MVI (P = 0.001) in the training set, and the model constructed using the radiomics features had good discriminative ability (AUC = 0.891) and calibration. However, radiomics analysis is complex and involves multiple steps. Wang et al[18] evaluated the methodological quality of radiomics models for predicting MVI; they found significant methodological heterogeneity and generally low quality, with an average radiomics quality score of 10 (out of 36). Moreover, most studies have focused solely on predicting the presence or absence of MVI in HCC. Therefore, additional data are needed to obtain valuable models for multilevel MVI stratification in HCC to meet the needs of individualized clinical evaluation. Our study not only predicted the MVI status of HCC but also further identified those at high risk (M2).

Given that HBV infection is the predominant risk factor for HCC in China, this study intended to target those with HBV-HCC. Compared to MRI, contrast-enhanced CT (CECT) is more affordable and clinically accessible. One study showed that radiomics models based on CT and MRI offer comparable predictive performance for MVI in solitary HCC[19]. Therefore, the main objective of this study was to develop prediction models that combine CECT radiomics and clinicoradiological factors to accurately identify MVI and M2 status in patients with HBV-HCC.

MATERIALS AND METHODS

Study population and design

Data from patients with positive hepatitis B surface antigen or anti-hepatitis B core antigen antibodies who underwent liver resection for histologically confirmed HCC were collected retrospectively at Anhui Provincial Hospital between January 1, 2020 and May 31, 2023. The study was approved by the Institutional Ethics Committee of our hospital (approval number: 2021-RE-043). Since this was a retrospective study, informed consent was waived.

Inclusion criteria were: (1) HCC classified as China liver cancer (CNLC) stages Ia, Ib, or IIa[20]; (2) Abdominal CECT performed within two preoperative weeks; and (3) No prior antitumor interventions before hepatectomy. Exclusion criteria included previous malignancies, incomplete clinical data, or poor image quality. Ultimately, 270 patients were included. Of these, 70% were randomly assigned to the training dataset and 30% to the validation dataset (Figure 1).

Figure 1 Flow diagram of the study enrolment patients. HBsAg: Hepatitis B surface antigen; HBcAb: Anti-hepatitis B core antigen antibodies; HCC: Hepatocellular carcinoma; CECT: Contrast-enhanced computed tomography.

Preoperative examination and histopathology

Demographic data and routine preoperative laboratory data, including liver function test results, routine blood test results, and α-fetoprotein (AFP) levels, were collected from the electronic medical records system. Preoperative diagnosis followed the criteria in the Guidelines for the Diagnosis and Treatment of Primary Liver Cancer (2022 edition)[20]. Two pathologists examined all surgical specimens using the "7-point" sampling method, focusing on detecting MVI. MVI is characterized by cancer cell clusters within blood vessels lined by endothelial cells, as observed microscopically. MVI was classified into three grades: M0 (no MVI detected), M1 (low-risk group, ≤ 5 MVIs in proximal non-neoplastic adjacent liver tissues), and M2 (high-risk group, > 5 MVIs in the proximal area or MVIs in distal non-neoplastic adjoining liver tissues)[21].

CT imaging protocol

Patients were scanned using the GE Discovery HD 750 multi-slice spiral CT scanner (GE Healthcare, Chicago, IL, United States). Initially, a plain abdominal CT scan was performed with the following parameters: Tube voltage of 120 kVp, tube current of 250-350 mA, slice thickness and spacing of 5 mm, field of view of 35-50 cm, matrix size of 512 × 512, rotation time of 0.7 seconds per revolution, and a pitch of 1.375. Following this, the ioversol contrast agent was injected via a high-pressure injector at a rate of 3.0 mL/second, with a dose of 1.5 mL/kg. The automatic scanning trigger software monitored the density of the descending aorta, initiating the arterial phase (AP) scan once it reached 95 Hounsfield units and started the scan 35 seconds later. Portal venous phase (PP) and delayed phase (DP) scans were performed 35 seconds and 3 minutes after the AP scan, respectively.

Qualitative analysis of CT images

Two surgeons (reader 1, with 3 years of experience, and reader 2, with 5 years of experience in liver CT reading) analyzed the CECT images retrospectively while blinded to clinical and pathological data. For patients with multiple tumors, only the largest one was analyzed. Discrepancies between the two readers were resolved through discussion until a consensus was reached. The following ten imaging features were evaluated based on previous studies: (1) Lobe involvement (0, one lobe; 1, multiple lobes); (2) Tumor margin (0, smooth; 1, non-smooth); (3) Tumor growth pattern (0, intrahepatic; 1, extrahepatic); (4) Enhancement pattern (1, typical with arterial hypervascularity and portal washout; 0, atypical); (5) Peritumoral arterial enhancement (0, absent; 1, present); (6) Internal arteries (0, absent; 1, present); (7) Intratumor necrosis (0, absent; 1, present); (8) Enhancing capsule (0, present; 1, absent); (9) Peritumoral hypointensity ring (0, present; 1, absent); and (10) Maximum tumor size[12,15,22].

Radiomics analysis of CT images

The radiomics analysis process involved tumor segmentation, feature extraction, feature selection, and model construction and evaluation (Figure 2).

Figure 2 Workflow of radiomics analysis. ROI^tumor: Region of interest of tumor; ROI^peri: Region of interest of 5 mm-wide peritumoral area; ICC: Intra-class correlation coefficient; ROC: Receiver operating characteristic; AUC: Area under the curve.

Tumor segmentation

AP, PP, and DP images were exported from the Picture Archiving and Communication Systems in DICOM format for each patient. An experienced surgeon (reader 3, with 5 years of liver CT reading experience) manually delineated the tumor region of interest (ROI^tumor) by tracing the tumor's edge layer-by-layer across the three-phase images using ITK-SNAP version 3.6 software (https://www.itksnap.org/). In cases with multiple lesions, only the most significant lesion was delineated. Moreover, readers 3 and 4 (5 years of CT experience) re-evaluated 25 randomly selected tumors one month later to ensure consistency. A 5-mm-wide peritumoral area (ROI^peri) was automatically generated using a dilation algorithm based on the tumor boundaries.

Radiomics feature extraction

We used linear interpolation to resample all images to an identical pixel size of 1 mm × 1 mm. Quantitative features were extracted from both the original and filtered images, including those transformed by techniques such as wavelet transformations, using the Pyradiomics package in Python (http://www.radiomics.io/pyradiomics.html). The source code of the radiomics analysis process can be found on the Github website (https://github.com/AIM-Harvard/pyradiomics). These features included first-order features, shape features, diagnostics features, gray level co-occurrence matrix, and gray level run length matrix. A total of 1515 features are extracted for each ROI^tumor and 1507 features for each ROI^peri (https://pyradiomics.readthedocs.io/en/latest/features.html). The extracted features were then standardized to a normal distribution with a mean of 0 and a SD of 1. All the radiomics parameters are described in Supplementary Table 1.

Feature selection and radscore

The reliability of the extracted radiomics features was calculated using the intra-class correlation coefficient analysis to ensure stability. Radiomics features with intra-rater and inter-rater correlation coefficients greater than 0.8 (indicating excellent stability) were selected for further investigation. Redundant features were eliminated through Pearson or Spearman’s rank correlation coefficient analysis. Feature reduction was performed using the Least Absolute Shrinkage and Selection Operator algorithm with tenfold cross-validation. Subsequently, the remaining radiomics features were integrated into radscore using the formula: Radscore = α + β₁X₁ + β₂X₂ + β₃X₃ + ... + β_nX_n (Formula 1). Here, α represents the constant term, X_n denotes the radiomics feature value, and β_n is the regression coefficient for each feature.

Construction and evaluation of the MVI status prediction model

Univariate analysis identified significant differences between patients who were MVI-positive and MVI-negative in the training dataset. Variables found substantial in this analysis were included in multivariate logistic regression to identify potential MVI risk factors. Prediction models were constructed using clinicoradiological risk factors, selected radiomics features, and combinations of these factors with the radscore. The training and validation datasets evaluated the models' classification performance through sensitivity, specificity, classification accuracy, receiver operating characteristic (ROC) curves, and AUC. Subsequently, paired ROC comparisons were conducted using the Delong test. A nomogram was created to visualize the clinicoradiological and radiomics features, enhancing interpretability. Calibration curves assessed the model’s fit[23], and the Hosmer-Lemeshow test evaluated the nomogram's calibration ability[24]. Decision curve analysis was performed to determine the clinical utility of the nomogram by measuring net benefits across various threshold probabilities for the entire cohort[25]. Lastly, two sensitivity analyses assessed how the prediction models changed in different settings. First, we tested the random forest algorithm as an alternative method to develop prediction models. Second, we rebuild the models specifically for patients over 50 years old.

Construction and evaluation of the M2 prediction model

A prediction model was developed to identify patients with M2, integrating clinicoradiological features and radscore. The methodology for this model mirrored that used for the MVI status prediction model. The model's classification performance, fit, clinical applicability, and sensitivity analyses were further evaluated.

Statistical analysis

Continuous variables are presented as means ± SD or medians with interquartile ranges, while categorical variables are reported as numbers (percentages). Comparisons of continuous variables were made using the unpaired Student’s t-test or Mann-Whitney U test, as appropriate. Categorical variables were compared using the χ² test or Fisher’s exact test. Model calibration was assessed using the Hosmer-Lemeshow goodness of fit test, and ROC curves among different prediction models were compared using the Delong test. Statistical analyses were conducted using SPSS (version 27, https://www.medcalc.org) and R software (version 4.3.2, http://www.rproject.org). A two-tailed P value < 0.05 was considered statistically significant.

RESULTS

Clinicoradiological characteristics

Table 1 presents the clinicoradiological characteristics of patients with varying MVI statuses. Univariate analysis revealed significant associations (P < 0.05) between MVI status and factors including neutrophil count, neutrophil-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, aspartate aminotransferase (AST), γ-Glutamyltransferase, total bilirubin, AFP, tumor diameter, CNLC stage, lobe involvement, tumor margin, internal arteries, intratumor necrosis, and peritumoral hypointensity ring. Multivariate analysis identified tumor diameter [odds ratio (OR) 1.434; 95%CI: 1.236-1.663; P < 0.001], non-smooth tumor margin (OR 3.469; 95%CI: 1.683-7.150; P < 0.001) and absence of peritumoral hypointensity ring (OR 7.521; 95%CI: 2.464-22.959; P < 0.001) as independent predictors of MVI. The AUCs for the combination of these three factors were 0.804 (95%CI: 0.741-0.868) in the training dataset and 0.774 (95%CI: 0.672-0.877) in the validation dataset. Supplementary Table 2 presents the clinicoradiological characteristics of patients with M2 status vs those with M0/M1. Figure 3 illustrates typical clinicoradiological features of patients with HCC who were MVI-positive and MVI-negative.

Figure 3 Two cases to show representative clinicoradiological factors of microvascular invasion-negative and microvascular invasion-positive hepatocellular carcinoma. A: A 63-year-old man presented with a 4.0 cm solid mass in the right lobe of the liver, featuring a peritumoral hypointensity ring (orange arrow) and smooth tumor margin; B: A 46-year-old man presented with a 15 cm solid mass in the right lobe of the liver, featuring a non-smooth tumor margin and the absence of a peritumoral hypointensity ring (yellow arrow).

Table 1 Comparisons of clinicoradiological factors for predicting microvascular invasion status in training and validation datasets, n (%).

Variables	Training dataset (n = 189)				Validation dataset (n = 81)
	MVI absent	MVI present	P value	Effect size	MVI absent	MVI present	P value
Sex			0.778				0.291
Female	10 (14)	13 (11)			2 (6)	7 (15)
Male	63 (86)	103 (89)			32 (94)	40 (85)
Age (years)			0.297				0.21
≤ 50	15 (21)	33 (28)			8 (24)	5 (11)
> 50	58 (79)	83 (72)			26 (76)	42 (89)
Alcohol			1				0.979
No	62 (85)	99 (85)			28 (82)	40 (85)
Yes	11 (15)	17 (15)			6 (18)	7 (15)
Liver cirrhosis			0.695				1
No	12 (16)	23 (20)			7 (21)	10 (21)
Yes	61 (84)	93 (80)			27 (79)	37 (79)
Antiviral therapy			0.377				0.073
No	43 (59)	77 (66)			14 (41)	30 (64)
Yes	30 (41)	39 (34)			20 (59)	17 (36)
Ascites			0.689				0.269
No	42 (58)	62 (53)			13 (38)	25 (53)
Yes	31 (42)	54 (47)			21 (62)	22 (47)
Neutrophil (10 × 9/L)	2.56 (2.1, 3.78)	2.98 (2.47, 3.98)	0.038	0.820	2.96 (2.2, 3.48)	3.1 (2.35, 3.76)	0.629
Lymphocyte (10 × 9/L)	1.44 (1.22, 1.78)	1.29 (0.98, 1.74)	0.09		1.47 ± 0.58	1.56 ± 0.51	0.514
Platelets (10 × 9/L)	143 (116, 179)	154 (111, 196.5)	0.368		141.5 (124.75, 182.25)	169 (99.5, 204.5)	0.897
Monocyte (10 × 9/L)	0.39 (0.28, 0.53)	0.4 (0.33, 0.53)	0.341		0.42 (0.33, 0.52)	0.42 (0.36, 0.58)	0.789
NLR	1.65 (1.37, 2.6)	2.37 (1.6, 3.56)	0.002	0.735	2 (1.43, 2.89)	1.87 (1.48, 2.46)	0.838
PLR	98.8 (75.56, 135.94)	115.55 (84.56, 157.42)	0.07		102.76 (81.42, 134.8)	101.33 (69.76, 129.07)	0.348
LMR	3.56 (2.79, 4.67)	3.12 (2.2, 4.36)	0.027	1.191	3.43 ± 1.29	3.58 ± 1.36	0.612
Alanine aminotransferase (U/L)			0.505				0.125
≤ 40	49 (67)	71 (61)			24 (71)	24 (51)
> 40	24 (33)	45 (39)			10 (29)	23 (49)
Aspartate aminotransferase (U/L)			0.001	0.249			0.102
≤ 35	47 (64)	45 (39)			23 (68)	22 (47)
> 35	26 (36)	71 (61)			11 (32)	25 (53)
γ-Glutamyltransferase (U/L)			0.009	0.200			0.02
≤ 45	40 (55)	40 (34)			19 (56)	13 (28)
> 45	33 (45)	76 (66)			15 (44)	34 (72)
Total bilirubin (μmol/L)			0.039	0.164			0.508
≤ 23	66 (90)	90 (78)			31 (91)	40 (85)
> 23	7 (10)	26 (22)			3 (9)	7 (15)
Albumin (g/L)			0.115				1
≥ 40	44 (60)	55 (47)			23 (68)	31 (66)
< 40	29 (40)	61 (53)			11 (32)	16 (34)
Prealbumin (g/L)			0.291				0.234
≥ 180	51 (70)	71 (61)			28 (82)	32 (68)
< 180	22 (30)	45 (39)			6 (18)	15 (32)
Alkaline phosphatase (U/L)			0.116				0.108
≤ 135	62 (85)	86 (74)			32 (94)	38 (81)
> 135	11 (15)	30 (26)			2 (6)	9 (19)
Prothrombin time (second)			0.843				1
≤ 14	62 (85)	101 (87)			31 (91)	43 (91)
> 14	11 (15)	15 (13)			3 (9)	4 (9)
International normalized ratio	1.07 (0.99, 1.11)	1.04 (0.98, 1.09)	0.21		1.06 (1, 1.12)	1.03 (0.98, 1.1)	0.271
HBV DNA load (IU/mL)			0.972				1
< 10000	62 (85)	97 (84)			26 (76)	36 (77)
≥ 10000	11 (15)	19 (16)			8 (24)	11 (23)
α-fetoprotein (μg/L)			0.008	0.206			0.061
≤ 7	37 (51)	35 (30)			18 (53)	14 (30)
> 7	36 (49)	81 (70)			16 (47)	33 (70)
Tumor diameter (cm)	4 (3, 6)	6.25 (4, 9.5)	< 0.001	0.590	4 (3, 5.88)	6 (3, 9.1)	0.027
Tumor number			0.375				0.26
Solitary	70 (96)	107 (92)			34 (100)	44 (94)
Multiple	3 (4)	9 (8)			0 (0)	3 (6)
CNLC stage			< 0.001	0.294			0.034
Ⅰa	49 (67)	43 (37)			24 (71)	21 (45)
Ⅰb	22 (30)	65 (56)			10 (29)	23 (49)
Ⅱa	2 (3)	8 (7)			0 (0)	3 (6)
Child-Pugh			0.994				1
A	60 (82)	94 (81)			30 (88)	42 (89)
B	13 (18)	22 (19)			4 (12)	5 (11)
Lobe involved			< 0.001	0.308			0.002
Solitary	72 (99)	88 (76)			34 (100)	36 (77)
Multiple	1 (1)	28 (24)			0 (0)	11 (23)
Tumor margin			< 0.001	0.337			0.008
Smooth	42 (58)	28 (24)			21 (62)	14 (30)
Non-smooth	31 (42)	88 (76)			13 (38)	33 (70)
Tumor growth pattern			0.945				1
Intrahepatic growth	64 (88)	100 (86)			29 (85)	41 (87)
Extrahepatic growth	9 (12)	16 (14)			5 (15)	6 (13)
Enhancement pattern			1				0.177
Atypical	23 (32)	37 (32)			10 (29)	22 (47)
Typical	50 (68)	79 (68)			24 (71)	25 (53)
Peritumoral arterial enhancement			0.444				1
Absent	67 (92)	101 (87)			30 (88)	41 (87)
Present	6 (8)	15 (13)			4 (12)	6 (13)
Internal arteries			< 0.001	0.33			0.008
Absent	43 (59)	30 (26)			24 (71)	18 (38)
Present	30 (41)	86 (74)			10 (29)	29 (62)
Intratumor necrosis			< 0.001	0.275			0.033
Absent	38 (52)	29 (25)			19 (56)	14 (30)
Present	35 (48)	87 (75)			15 (44)	33 (70)
Enhancing capsule			0.236				0.158
Absent	40 (55)	52 (45)			17 (50)	15 (32)
Present	33 (45)	64 (55)			17 (50)	32 (68)
Peritumoral hypointensity ring			0.001	0.250			0.115
Absent	20 (27)	10 (9)			10 (29)	6 (13)
Present	53 (73)	106 (91)			24 (71)	41 (87)

Continuous variables were expressed as means ± SD or medians with interquartile ranges. Both antiviral therapy status and hepatitis B virus DNA load were collected from patients on admission. MVI: Microvascular invasion; NLR: Neutrophil-to-lymphocyte ratio; PLR: Platelets-to-lymphocyte ratio; LMR: Lymphocyte-to-monocyte ratio; CNLC stage: China liver cancer stage; HBV: Hepatitis B virus.

Radiomics analysis

Radiomics analysis used six feature subsets related to MVI (Supplementary Table 3) derived from 6 ROIs. Subsequently, six radiomics models were constructed based on these subsets. The radscores for each patient were calculated using Formula 1 from the materials and methods section. The feature subsets and their corresponding radscores were defined as follows: ROI^tumor-AP represents the feature subset based on the AP tumor ROI, with the radiomics score denoted as radscore (ROI^tumor-AP). ROI^peri-AP represents the feature subset based on the AP 5 mm-wide peritumoral area ROI, with the radiomics score denoted as radscore (ROI^peri-AP). The remaining feature subsets and radscores follow a similar naming convention. In the training dataset, the radiomics models exhibited sensitivities between 46.6% and 83.6%, specificities between 68.5% and 96.0%, accuracies between 64.0% and 81.0%, and AUC values between 69.0% and 87.0% for MVI classification. In the validation dataset, sensitivities ranged from 40.4% to 78.7%, specificities from 61.8% to 82.4%, accuracies from 56.8% to 72.8%, and AUC values from 59.4% to 72.0% for MVI classification. The ROI^tumor-DP-based model achieved the highest classification performance for MVI, with an AUC of 87.0% (95%CI: 81.9%-92.2%) in the training dataset and 69.6% (95%CI: 57.9%-81.3%) in the validation dataset (Supplementary Table 4). To further assess the classification performance of the radiomics models, pairwise comparisons of ROC curves were conducted using the Delong test, with the ROI^tumor-DP model as the reference. The ROI^tumor-DP model exhibited significantly better performance in the training dataset than the other four models, except the ROI^peri-AP model (P = 0.075; Supplementary Table 5). However, no significant differences were observed in the validation dataset. Univariate analysis indicated that all six radscores were significant risk factors for MVI status (P < 0.001). Similarly, six feature subsets associated with M2 were identified, and their corresponding radiomics scores were computed (Supplementary Table 6). The univariate analysis indicated that all six radscores were significant risk factors for M2 (P < 0.001).

Construction and evaluation of the MVI status prediction model

Univariate analysis identified 20 significant predictors among all clinicoradiological and radiomics factors, including seven clinical variables, seven imaging features, and six radscores. The multivariate regression analysis pinpointed three independent risk factors for MVI status: Non-smooth tumor margin, absence of a peritumoral hypointensity ring, and a higher radscore based on ROI^tumor-DP. These factors were integrated into the radiologic-radiomics (RR) model, expressed by the formula: Y = -2.346 + 1.679 × peritumoral hypointensity ring + 1.880 × radscore (ROI^tumor-DP) + 0.976 × tumor margin (Table 2). The resulting RR model demonstrated robust performance in predicting MVI risk, with an AUC of 0.841 (95%CI: 0.783-0.898) in the training dataset and 0.768 (95%CI: 0.664-0.872) in the validation dataset (Figure 4A). An optimal threshold was determined by maximizing the Youden index from ROC analysis conducted with the training dataset. At this threshold, the model achieved a sensitivity of 88.8%, specificity of 64.4%, and accuracy of 79.4% in the training dataset. In the validation dataset, sensitivity was 80.9%, specificity 61.8%, and accuracy 72.8%. The calibration curve indicated excellent alignment between predicted and observed MVI status for both datasets (Figure 4B and C). The Hosmer-Lemeshow test yielded χ² values of 8.327 (P = 0.402) for the training dataset and 12.804 (P = 0.119) for the validation dataset, suggesting a good model fit. Further analysis demonstrated that this model for MVI prediction provided a significantly higher net benefit than scenarios with universal intervention or no intervention when the threshold probability exceeded 0.12 (Figure 4D). The RR model is illustrated as a nomogram in Figure 4E. However, its performance was comparable to models based solely on clinicoradiological factors (AUC 0.841 vs 0.804; P = 0.095), and ROI^tumor-DP radiomics features (AUC 0.841 vs 0.870; P = 0.143) in the training dataset. Similarly, in the validation dataset, the performance of the predictive model was comparable to models based on clinicoradiological factors (AUC 0.768 vs 0.774; P = 0.850) and ROI^tumor-DP radiomics features (AUC 0.768 vs 0.696; P = 0.120).

Figure 4 Calibration, discrimination, and clinical utility analysis of the radiologic-radiomics model for predicting microvascular invasion. A: Receiver operating characteristic analysis curve of the model; B-D: Calibration curves in the training (B) and validation (C) datasets, and decision curve (D) in the overall patients; E: The model is presented with a nomogram scaled by the proportional regression coefficient of each predictor. AFP: Alpha-fetoprotein; MVI: Microvascular invasion; radscore (ROI^tumor-DP): Represents the radscore based on regions of interest of tumor from delayed phase contrast-enhanced computed tomography images; ROC: Receiver operating characteristic; AUC: Area under the curve.

Table 2 Multivariate analyses of clinicoradiological factors and radiomics for predicting microvascular invasion status in the training dataset.

Factors	Coefficient	P value	OR	95%CI
Tumor margin	0.976	0.012	2.655	1.241-5.679
Peritumoral hypointensity ring	1.679	0.002	5.360	1.857-15.477
Radscore (ROItumor-DP)	1.880	< 0.001	6.552	3.364-12.761
Intercept	-2.346	< 0.001	0.096

Radscore (ROI^tumor-DP): Radscore based on regions of interest of tumor from delayed phase contrast-enhanced computed tomography images: OR: Odds ratio.

Construction and validation of the M2 prediction model

Univariate analysis of 37 clinicoradiological factors identified 18 variables significantly associated with M2 (P < 0.05; Supplementary Table 2). Subsequent multivariate logistic regression analysis on these 18 clinicoradiological factors, along with six radscores, revealed that AFP level (OR 3.818; 95%CI: 1.380-10.563; P = 0.01), enhancing capsule (OR 3.962; 95%CI: 1.628-9.643; P = 0.002), AST level (OR 3.760; 95%CI: 1.485-9.520; P = 0.005), and radscore (ROI^peri-AP) (OR 5.967; 95%CI: 2.609-13.649; P < 0.001) were independent predictive factors for M2. An M2 risk prediction model was constructed using a linear equation incorporating these four factors. In the training dataset, the model demonstrated a sensitivity of 78.6%, specificity of 86.4%, accuracy of 84.7%, and an AUC of 86.5% (95%CI: 79.7%-93.4%). In the validation dataset, sensitivity was 57.1%, specificity 93.3%, accuracy 84.0%, and AUC 79.8% (95%CI: 67.6%-91.9%) (Figure 5A). The calibration curve strongly agreed between the model’s predicted probabilities and M2 outcomes in both sets (Figure 5B and C). The Hosmer-Lemeshow test further confirmed good calibration, with χ² values of 9.311 for the training dataset (P = 0.317) and 2.280 for the validation dataset (P = 0.131). Clinical utility was assessed using decision curve analysis across a cohort of 270 patients with HCC. Figure 5D illustrates the decision curve analysis, indicating that the M2 prediction yielded higher overall patient net benefits than universal or no intervention scenarios when the threshold probability exceeded 5%. A nomogram based on the M2 prediction model is presented in Figure 5E.

Figure 5 Calibration, discrimination, and clinical utility analysis of the model for predicting high-risk group. A: Receiver operating characteristic analysis curve of the model; B-D: Calibration curves in the training (B) and validation (C) datasets, and decision curve (D) in the overall patients; E: The model is presented with a nomogram scaled by the proportional regression coefficient of each predictor. AFP: Alpha-fetoprotein; AST: Aspartate aminotransferase; MVI: Microvascular invasion; radscore (ROI^peri-AP): Represents the radscore based on regions of interest of 5 mm-wide peritumoral area from arterial phase contrast-enhanced computed tomography images; M2: High-risk group; ROC: Receiver operating characteristic; AUC: Area under the curve.

Sensitivity analysis

First, we used the random forest algorithm as an alternative method to develop prediction models. The results showed that the MVI prediction model yielded an AUC of 0.761 (95%CI: 0.656-0.865), and the M2 prediction model yielded an AUC of 0.759 (95%CI: 0.644-0.874) in the validation dataset. This means that using the random forest method did not improve the AUC compared to the stepwise procedure method. Then, we rebuilt a new MVI prediction model based on patients over 50 years of age, which resulted in an AUC of 0.867 (95%CI: 0.811-0.923) in the training data and an AUC of 0.662 (95%CI: 0.517-0.806) in the validation data. Furthermore, all variables in the new prediction model were included in the abovementioned MVI status prediction model. Similarly, a new M2 prediction model was created, resulting in an AUC of 0.877 (95%CI: 0.792-0.961) in the training dataset and an AUC of 0.700 (95%CI: 0.549-0.848) in the validation dataset. Two-thirds of the variables in the model were included in the original M2 prediction model.

DISCUSSION

Accurate preoperative assessment of MVI is essential for developing personalized treatment plans and clinical management strategies for patients with HBV-HCC. This study produced two predictive models: One for preoperative evaluation of MVI presence and another to predict M2 status in patients with HBV-HCC. Both models demonstrated strong classification performance, calibration, and clinical utility. The different sensitivity analyses showed similar AUC.

Zhang et al[26] developed an MVI status classifier that integrated radiomics signature and two clinical factors (age and AFP level) and achieved AUC of 0.806, 0.803, and 0.796 in the training, test, and independent validation cohorts, respectively. For MVI risk stratification (low risk vs high risk), the AUCs of the model were 0.783, 0.778, and 0.740, respectively. Another study developed a model of radiomics and clinical data (age and AFP), which achieved an AUC of 0.85 and 0.77 in the training and test groups, respectively[27]. In conclusion, through comparison and analysis with existing models, our model is roughly the same as theirs in terms of performance.

Our analysis of radiomics features demonstrated that texture features have substantial predictive value for MVI, consistent with previous studies[26,28-30]. Xu et al's research[22] revealed that radiomics features related to tumor size (e.g., shape surface area) were the most critical components for predicting histologic MVI, and Peng et al's research[28] also yielded similar results. However, contrary to these two studies, no association between shape features and MVI was identified. In our study, tumor boundaries were manually delineated, whereas the above two studies used a semi-automated method to obtain the ROI^tumor. This discrepancy may be due to the subjective nature of manual tumor delineation, where continuous observation and image analysis can induce visual fatigue and reduced accuracy. Consequently, some relevant shape features may have been excluded during intra-class correlation coefficient analysis. Therefore, developing automated tumor delineation algorithms remains critical for future research.

In the MVI status prediction model, the radscore (ROI^tumor-DP) (OR 6.552) was a stronger predictor than tumor margin (OR 2.655) and the absence of a peritumoral hypointensity ring (OR 5.36). Similarly, in the M2 status prediction model, the radscore (ROI^peri-AP) (OR 5.967) was more critical than the AFP level (OR 3.818), AST level (OR 3.760), and enhancing capsule (OR 3.962). These findings are consistent with those of Peng et al[28] who identified radiomics signatures as the primary predictors, followed by clinicoradiological factors. Based on the Delong test, our results showed no significant difference in MVI predictive ability among the RR, radiomics-only, and clinicoradiological models. This observation suggests that including radiomics features did not enhance the prediction model’s performance. Similarly, a study by Xu et al[22] demonstrated that combining clinicoradiological and radiomics features produced satisfactory results for predicting MVI. However, radiomics did not significantly improve the predictive value beyond traditional radiological factors. Furthermore, Meng et al's study[31] indicated that radiomics outperformed clinical indicators predicting MVI, specifically in HCC tumors measuring 2.0-3.0 cm in diameter. However, for tumors measuring 3.1-4.0 cm and 4.1-5.0 cm, the predictive performance of both models was comparable. In summary, while radiomics offers a high-throughput approach, it does not necessarily outperform traditional clinical models in predicting MVI in HCC. Future research should aim to clarify the specific populations that could benefit most from radiomics-based models.

Previous studies have demonstrated that AFP levels play a significant role in predicting MVI preoperatively, a finding consistent with our results[32-34]. However, the threshold for AFP in predicting MVI remains debated. Our study used AFP levels > 7 µg/L (based on our hospital laboratory’s reference value of ≤ 7 µg/L) as the threshold for predicting M2. In contrast, Xu et al[22] identified preoperative AFP levels > 400 µg/L as predictive of MVI, and a retrospective analysis by He et al[35] reported that AFP > 158 µg/L was associated with an increased risk of MVI. Therefore, multi-center studies with larger sample sizes are necessary to better evaluate the predictive utility of AFP. AST, a key liver enzyme, is commonly employed for assessing liver inflammation. Elevated AST levels signify severe liver inflammation, which can contribute to carcinogenesis and tumor metastasis[36]. Numerous studies have established a correlation between MVI and AST levels[37,38]. In our study, patients with M2 status had higher rates of abnormal AST levels than those with M0 and M1 status.

The relationship between an enhancing capsule and MVI remains contentious. Many researchers believe that HCC with an incomplete enhanced capsule is more susceptible to developing MVI[39,40]. In this study, the M2 group had a higher proportion of tumors with incomplete enhancing capsules compared to the non-MVI group. Multivariate analysis showed that an incomplete enhancing capsule was a risk factor for M2 (OR 3.691, P = 0.004). This finding could be attributed to the protective function of the capsule, as its incompleteness likely increases the risk of cancer cells invading blood vessels. Segal et al[41] identified specific CT imaging features, such as “internal arteries” and “hypodense halos”, as having predictive solid values for MVI. Similarly, our study confirmed that a peritumoral hypointensity ring is strongly associated with MVI. Previous research has shown that a non-smooth tumor margin indicates a higher MVI risk[22]. Our findings further support this, identifying a non-smooth tumor margin as an independent risk factor for MVI. Although tumor morphology can aid in the preoperative prediction of MVI, its assessment frequently depends on subjective clinical judgment and lacks objective quantitative metrics.

HBV infection is a significant risk factor for liver cirrhosis and HCC. The HBV X protein is a critical regulatory multifunctional protein of the virus linked to vascular invasion[42]. One study reported that a high preoperative HBV-DNA load (> 10⁴ IU/mL) is a risk factor for MVI in patients with HCC[12]. In our study, most patients had a low HBV-DNA load (< 10⁴ IU/mL), which may explain why viral load was not identified as a predictive factor. HBV genotype C, prevalent in mainland China (68.3%), promotes chronic infection[43,44]. Moreover, genetic susceptibility in the Chinese population contributes to persistent inflammatory responses that hinder the complete clearance of HBV, leading to chronic infection and long-term viral replication[45,46]. The development of MVI is a gradual process, and the HBV-DNA levels at the time of admission may not accurately capture the relationship between viral load and MVI.

This study has several limitations. First, it was a retrospective, single-center study with a small sample size. More extensive, multi-center randomized clinical trials are necessary to validate the utility of our models in guiding personalized surgical plans and clinical management for patients with HBV-HCC. Second, manual segmentation was employed to delineate the regions of interest susceptible to subjective bias and may introduce variability in results. As artificial intelligence-assisted imaging algorithms continue to advance, fully automated MVI prediction could enhance the accuracy and reproducibility of this approach.

CONCLUSION

Our models, incorporating CECT radiomics signatures and clinicoradiological factors, show considerable predictive accuracy for MVI and M2 status in patients with HBV-HCC and hold promise for guiding personalized surgical planning and clinical management.

ACKNOWLEDGEMENTS

The authors thank Dr. Yu Hu, Dr. Hao Chen, and Dr. Yu Fang (Department of General Surgery, Anhui Provincial Hospital, The First Affiliated Hospital of USTC) for their invaluable support in CECT image analysis.