Imaging and Radiomics Study of Microvascular Inltration of Primary Liver Cancer Using a Seven-point Pathological Sampling Method

Purpose (cid:0) In this study, the aim was to assess the imaging features and radiomics of microvascular inltration (MVI) of primary liver cancer (PLC) under the control of a seven-point pathological sampling method. Methods: The data of 37 patients with PLC who underwent surgical resection in our hospital from October 2018 to September 2019 were retrospectively collected. Postoperative pathological specimens were collected using a seven-point sampling method to determine the presence of MVI. Preoperative CT and MRI scans were performed to characterize the tumors. Findings from the imaging studies were imported into the radiomics platform, and 70% and 30% of the data were randomly assigned to the training and validation sets, respectively. Lastly, support vector machine (SVM) classiers were used to classify liver lesions into their respective pathological types. Results: Differences in tumor morphology and satellite lesions were statistically signicant between the MVI positive and MVI negative groups on CT images. On MRI, there were statistically signicant differences between the MVI positive and MVI negative groups in peripheral enhancement of the arterial phase (AP) and peripheral low signal in the hepatobiliary phase (HBP). In the radiomics analysis, the imaging features extracted from the AP had strong predictive power in both groups (CT and MRI). For the phase images, 15 and 12 valuable features from CT and MRI were selected to develop the radiomics signature, respectively. The AUCs of the training set were 0.965 (sensitivity: 0.979; specicity: 0.931; precision: 0.939) and 0.962 (sensitivity: 0.963; specicity: 0.897; precision: 0.923) , the validation set were 0.842 (sensitivity: 0.967; specicity: 0.733; precision: 0.714) and 0.769 (sensitivity: 0.846; specicity: 0.727; precision: 0.727). The PVP also performed well on CT (AUC: 0.851/0.891) and MRI (AUC: 0.886/0.846). The predictive power was not enhanced by combining the features of multi-phase images. Conclusions: This was a controlled study on preoperative CT and MRI imaging and radiomics based on a seven-point pathological sampling method can avoid false-negative results caused by traditional pathological sampling. The imaging analysis results obtained and the radiomics prediction model established in this study may be more accurate than conventional models. AP, arterial phase;


Introduction
PLC is one of the most common malignant tumors in the world. Due to substantial advances in diagnostic imaging in recent years, the early detection rates of PLC has increased globally, leading to improved survival rates. Currently, resection of the primary tumor and liver transplantation are the two primary surgical strategies used for PLC, yet the 5-year recurrence rate after surgery is high at 35-70% [1]. MVI is an essential factor that affects the postoperative recurrence and prognosis of cancer patients [2], including those with PLC [3,4]. Microvascular in ltration refers to the presence of tumor cells in the vascular space lined by endothelial cells, which primarily consists of portal vein branches (including intracapsular vessels) [5].
At present, the gold standard for diagnosing PLC MVI is observation via postoperative pathological microscopy [6]. The periphery of PLC tumors is the representative region of biological tumor behavior [7].
According to the 2017 edition of the guidelines for the diagnosis and treatment of PLC, for a single tumor with a maximum diameter of > 3 cm, the seven-point baseline sampling method should be used for the pathological assessment of gross specimens [8]. In a previous retrospective analysis on preoperative imaging of MVI, the seven-point sampling method and image location were not compared when using the pathological gold standard [9,10]. However, some patients who actually have MVI may have falsenegative results due to the lack of pathologically-determined MVI, leading to inaccurate ndings.
In this study, image control was performed on postoperative pathology using a seven-points sampling method to accurately analyze the preoperative imaging manifestations of PLC MVI and to conduct radiomics modeling.

Clinical data
The data from 37 patients diagnosed with PLC who underwent surgical resection in our hospital from October 2018 to September 2019 were collected. The inclusion criteria included the following: a single primary lesion in the liver with a diameter of > 3 cm, as examined by CT and MRI within one month before surgery, along with PLC determined by postoperative pathology. The exclusion condition was a single primary lesion diameter of ≤ 3 cm. Preoperative interventional examinations were performed, and the lesions were multiple intrahepatic diffuse PLC. In this experiment, hepatobiliary surgeons performed intraoperative ligation and localization of lesion specimens. In Fig. 1a, one suture line represented the front of the human body, and two sutures represented the right side of the human body. The largest section of the lesion was incised along the transverse plane of the body (Fig. 1a).

Pathological materials
Postoperative gross specimens of all enrolled patients were collected using the seven-point baseline sampling method at which 12, 3, 6, and 9 o'clock positions of the maximum section of the tumor were sampled. The samples were collected at the junction of the carcinoma and para-carcinoma liver tissue at 1:1 (A, B, C, and D); at least one piece was taken from inside the tumor (E); and one piece of liver tissue was collected from the tumor margins ≤ 1 cm (near the tumor) and > 1 cm (far near the tumor) (F,G), as shown in Fig. 1b. The pathological diagnosis was reported according to the seven points of sampling, which were jointly completed by two pathologists with more than 12 years of combined work experience. According to the postoperative pathological ndings, 37 patients were divided into the MVI positive group and MVI negative group. The patients with MVI at one or more of the sampling points were de ned as the MVI positive group, while those without MVI were de ned as the MVI negative group. Since MVI tends to occur at the junction between the tumor and healthy liver tissues, this study only conducted statistical analyses on the results of A, B, C, and D.
Imaging data CT: Thirty-three patients underwent enhanced upper abdominal CT. Twenty cases were scanned using the Siemens dual-source spiral CT (SOMATOM De nition Flash, Germany), and the scanning parameters were: thickness of 1.5 mm, pitch of 0.8, tube voltage of 120 kV, and tube current of 210 mAs. The GE 64slice spiral CT scanner (Light Speed VCT, Germany) was used in 13 cases. The scanning parameters were as follows: thickness of 1.25 mm, pitch of 0.984, tube voltage of 100 kV, tube current of 400 mAs. First, the patients underwent a scan of the whole liver, followed by enhanced scanning. A contrast agent was used in the imaging sessions (Iopromide, 300 mgI/mL, Germany Bayer Healthcare Co.), with a dosage of 1.5 mL/kg and blood ow rate of 3 mL/s through the super cial vein on the elbow with a double-barrel high-pressure syringe. At 30 s, 60 s, and 120 s post-injection, the arterial phase, portal vein phase, and delayed phase were scanned, respectively. According to the pathological results, 33 patients were divided into the MVI positive group and MVI negative group. The size, shape (round/irregular), and presence of satellite lesions of the tumor were analyzed by comparing the pathological specimens of the patients with the CT images.
MRI: Twenty-three patients underwent enhanced MR imaging with 10 mL Gd-EOB-DTPA at 0.25 mmol/mL (Germany Bayer Healthcare Co.) using the Siemens Verio 3.0 T MRI scanner (Germany), with a 12channel body phased-array coil. Gd-EOB-DTPA was administered as a bolus injected at a rate of 1 mL/s through the cubital vein followed by a 20 mL saline chaser administered at the same rate. The scanning parameters of T1WI volume interpolated body examinations (VIBE) included the following: repetition time (TR) of 3.9 ms, echo time (TE) of 1.4 ms, ip angle of 9° (30° in hepatobiliary phase), eld of view (FOV) of 350 mm, matrix of 168 × 320, voxel size of 1.6 × 1.1 × 4.5 mm, signal to noise ratio (SNR) of 1.00, and section thickness of 4.5 mm. The scanning parameters of T2WI using the BLADE technique were TR 2930 ms, TE 189 ms, FOV 400 mm, voxel size 1.3 × 1.3 × 6.0 mm, SNR 1.00, and 6 mm section thickness. The scanning parameters of diffusion-weighted imaging (DWI) were as follows: TR of 9000 ms, TE of 66 ms, FOV of 420 mm, matrix of 80 × 148, voxel size of 3.5 × 2.8 × 6.0 mm, SNR of 1.00, and section thickness of 6 mm. Delay phase scanning was performed at 5, 10, and 20 min after the administration of Gd-EOB-DTPA.
The points (A, B, C, and D) of 23 patients were divided into the MVI positive group and MVI negative group, based on the pathological ndings. Next, the corresponding points on the MRI images were compared, and the tumor boundaries (clear/fuzzy), peripheral enhancement in AP, and peripheral low signal in HBP were analyzed.

Image segmentation
A total of 13 patients who were examined by CT before the operation and con rmed by pathology after the operation to have MVI were selected for the radiomics analysis, along with ve patients who were examined by the same imaging tool before surgery and con rmed by pathology to have no MVI were randomly selected. The points (A, B, C, and D) of 18 patients (72 points in total) were selected for image segmentation under pathological control. Eleven patients, who had undergone MRI examinations before surgery and were con rmed to have MVI by post-surgery pathology, were selected. Four patients who had been examined by the same imaging tool before surgery and were con rmed to have no MVI were randomly selected. The points (A, B, C, and D) of 15 patients (60 points in total) were selected for image segmentation under pathological control.
The regions of interest (ROI) were drawn in the cross-section of the corresponding images of the pathological specimens, using a rectangle similar to the shape of the pathological specimen. The ROI was selected from the AP, portal venous phase (PVP), and delayed phase (DP) of CT images (MRI added the HBP), and was delineated by a junior radiologist who had two years of work experience. Each segmentation was re-examined by a senior radiologist with more than 11 years of work experience (Figs. 2 and 3).

Radiomic feature extraction and machine learning
This study included 18 CT cases and 15 MRI cases. The former included arterial, venous, and delayed phase images, while the latter included more hepatobiliary phase images. All images were transferred into the radiomics platform of Huiying Medical Technology (Beijing, China). Radiomic features were subdivided into rst-order, shape, texture, and lter-based features, accounting for 1,395 in total. The features were extracted from liver lesions outlined in all sections of each phase image. Only a small number of patients included in this study developed tumor invasion. Therefore, before dimensionality reduction of the data was performed, the sample equilibrium was carried out using the SMOTE (Synthetic Minority Oversampling Technique) algorithm. The Variance Threshold, Select K Best, and Lasso methods were used to reduce the dimensionality of features and lter out the imaging features with high diagnostic e ciency in each phase. In total, 70% of the data were randomly assigned to the training set, while the other 30% to the validation set. In addition, support vector machine (SVM) classi ers were used to classify liver lesions into the respective pathological types.

Statistical analysis
The prediction performance of the SVM in each phase image was evaluated by receiver operating characteristic (ROC) curves of the training set, as well as the validation set. Sensitivity, speci city, precision, and 95% CI were calculated to diagnose the SVM model. SPSS version 22.0 software (IBM, Chicago, IL, USA) was used for statistical analysis. Between-group comparisons were conducted with the chi-squared test for categorical variables and the independent sample t-test for continuous variables. Multivariate analysis was performed using logistic regression. Signi cance was set at p < 0.05.

Clinical features
In total, 37 patients (30 males and seven females) aged from 33 to 74 years old were enrolled in this study, including 35 patients with hepatocellular carcinoma (HCC) and two patients with combined hepatocellular carcinoma and cholangiocarcinoma (cHCC-CC). Postoperative pathology was used to con rm 16 cases (24 points in total) in the MVI positive group and 21 cases in the MVI negative group. Gender, age, hepatitis b surface antigen (HBsAg), alpha-fetoprotein (AFP), des-γ-carboxy Prothrombin (DCP), and liver cirrhosis were compared among the two groups, and the differences were not statistically signi cant (Table 1).

Imaging analysis
In this study, 33 patients underwent preoperative CT examinations. There were 15 cases in the MVI positive group, and 14 cases with the maximum tumor diameter of ≥ 5 cm. The tumors were irregularly shaped in 10 cases. Satellite lesions were observed around the tumor in eight of the cases. There were 18 cases in the MVI negative group, and 12 cases with the maximum tumor diameter ≥ 5 cm. The tumors were irregularly shaped in four of the cases. Satellite lesions were observed around the tumor in one case (Table 2). Univariate analysis revealed that the difference between the two groups was statistically signi cant in the two factors of irregular tumor morphology and satellite lesions around the tumor. The multivariate logistic regression analysis showed that the model was statistically signi cant (P = 0.010). OR, odds ratio; 95% CI, 95% con dence interval; MVI, microvascular in ltration. *P < 0.05 was statistically signi cant.
Next, 23 patients underwent preoperative MRI examinations. In the MVI positive group, there were 12 cases, with an MVI total of 24/48 points, including 20 points with a fuzzy boundary. There were 11 points of enhancement around the tumor in the AP and 16 points of low signal around the tumor in the HBP. In the MVI negative group, there were 11 cases with a total of 44 points, including six points with a fuzzy boundary. There were two points of enhancement around the tumor in the AP and four points of low signal around the tumor in the HBP (Table 3). Univariate analysis showed that the differences between the three factors, including fuzzy tumor boundary, peripheral enhancement in the AP, and peripheral low signal in the HBP, were statistically signi cant between the two groups. The multivariate logistic regression analysis showed that the model was statistically signi cant (P = 0.017). OR, odds ratio; 95% CI, 95% con dence interval; AP, arterial phase; HBP, arterial phase; MVI, microvascular in ltration. *P < 0.05 was statistically signi cant.

Machine learning
In total, 15, 11, and 15 kinds of effective features were extracted from the AP, PVP, and DP of the enhanced CT images, which were correlated with MVI. In addition, 12, 12, 13, and 9 kinds of effective features were extracted from the AP, PVP, DP, and HBP of MRI images, which were associated with MVI (Tables 4 and 5). The SVM classi er was used for machine learning, and the ROC curve of the model established in each phase, along with the multi-phase of CT and MRI, are shown in Fig. 4. The evaluation indexes of each phase model are shown in Table 6.   AUC, area under the receiver-operating characteristic curve; 95% CI, 95% con dence interval; AP, arterial phase; PVP, portal venous phase; DP, delayed phase; HBP, hepatobiliary phase; ALL, AP + PVP + DP + HBP; SVM, support vector machine; ROC, receiver-operating characteristic.
The imaging features extracted from the AP had strong predictive power in both the CT and MRI groups.

Discussion
Previous studies have shown that MVI is a major prognostic factor for PLC. For example, Sumie et al. found that the degree of MVI is related to the risk of postoperative disease recurrence, and the 2-year tumor-free survival rates of patients without MVI, mild MVI, and severe MVI after hepatectomy were 75.9%, 47.2%, and 32.7%, respectively [11]. Some studies have shown that certain factors affect whether PLC is associated with MVI, including tumor diameter, degree of tumor cell differentiation, AFP level, and the presence of more than one tumor [12]. Proteins induced by the absence of vitamin K or antagonist-(PIVKA-), AFP, alpha-fetoprotein lens culinaris agglutinin 3 (AFP-L3), and γ-glutamyl transpeptidase (GGT), are closely related to MVI [13]. Li et al. believed that among HCC patients with histologicallycon rmed MVI, the prognosis of patients under 60 years of age was worse than those over 60 years of age [14]. In this study, gender, age, HBsAg, AFP, DCP, and liver cirrhosis status were included as observational indicators. However, there were no statistically signi cant differences in these indicators between the MVI positive and MVI negative groups. It is inconsistent with the conclusions of previous studies on the correlation of MVI with AFP and age, which may be related to the small sample size. However, in the MVI positive group, the average age younger than the MVI negative group.
There is a hope that the early prediction of MVI can guide the clinical development of individualized treatment plans. Previously, Kim et al. reported that tumor size is correlated with MVI in univariate analyses, but not in multivariate analyses [15]. Some scholars believe that PLC with a diameter of > 5c m is more likely to spread through MVI [16]. Ahn et al. reported that Gd-EOB-DTPA is an important predictor of MVI with tumor enhancement in AP images [17]. Similarly, periarterial enhancement may re ect the effect of MVI on hemodynamics in the peripheral PLC. While normal liver tissue supplies blood to the portal vein, MVI can lead to tumor thrombi formation. In return, small branched embolization of the portal vein may occur around the tumor, followed by a low-perfusion state and arterial hyper-perfusion compensation [18]. In the HBP after Gd-EOB-DTPA enhancement, the contrast agent available to normal hepatocytes showed high signal, while the tumor tissue showed low signal because it did not contain normal hepatocytes and could not absorb the contrast agent. Some scholars have suggested that weak peripheral signal of the hepatobiliary tumor could predict MVI more accurately, as the sensitivity and speci city are higher [15]. In another study, Lee et al. speculated that MVI might induce changes in blood perfusion around the tumor, affecting the function of organic anion transport peptides on the liver cell membrane, which is known to introduce Gd-EOB-DTPA into the liver cells. The abnormal function of the translocation polypeptide could reduce the uptake of Gd-EOB-DTPA in the peritumor hepatocytes, leading to a relatively low signal [18].
The tumor size, morphology, and presence of satellite lesions were analyzed on the preoperative enhanced CT images to predict the occurrence of MVI in our study. We found no signi cant difference in the possibility of MVI, regardless of whether the lesion diameter was greater than 5 cm. Our nding differs from previous studies, which may be related to the small sample size of this study. Meanwhile, we found that the irregular shape of the tumor and the satellite lesions around the tumor were related to the occurrence of MVI. Our study analyzed the differences between the MVI positive group and MVI negative group in fuzzy tumor margins, enhancement around the tumor in the AP, and low signal around the tumor in the HBP on preoperative MRI Gd-EOB-DTPA-enhanced images. In the univariate analysis, the three observation indicators were all related to MVI, which is consistent with previous studies. However, this study's innovation lies in the point-to-point analysis of the seven-point sampling method and MRI analysis. Our conclusions may be more accurate and allow for the avoidance of false-negatives caused by traditional sampling methods.
As a newer eld of research, radiomics combines medicine and engineering to convert image information into texture features for quantitative and digital research. In a previous study, Wilson et al. extracted texture features from T1, T2, AP, and PVP images of preoperative MRI from patients with HCC. Both tumor entropy and mean were found to be associated with MVI. The texture analysis of preoperative imaging correlated with microscopic features of HCC and can be used to predict patients with high-risk tumors [19]. Another study applied radiomics to analyze the preoperative CT enhanced images of 206 patients with HCC. Different dimensionality-reduction methods and feature classi ers were used for machine learning, and the e cacies of each model were compared. The models established with the LASSO + GBDT method showed optimal diagnostic performance and the greatest diagnostic value for MVI. Hence, radiomics can be used for the preoperative and noninvasive diagnoses of MVI, yet different dimensionality reduction and modeling methods will affect the nal model [20]. In another study, Yang

Limitations
The sample size of this study was small. Still, we believe that the results obtained were highly accurate with the cooperation of intraoperative positioning by clinicians and postoperative sampling by pathologists. In return, we believe that our study has prevented the occurrence of false positives. In future work, we plan to improve our multidisciplinary cooperation further, while also expanding our sample sizes.

Conclusions
This was a retrospective study on preoperative imaging (CT and MRI) and radiomics, based on a sevenpoint pathological sampling method, which can avoid false-negative results commonly associated with conventional pathological sampling methods. In return, the imaging analysis results and the established radiomics prediction model may be more accurate. In the preoperative CT imaging features, irregular morphology, and satellite lesions around the tumor were found to be associated with the occurrence of MVI. In terms of preoperative MRI, MVI was correlated with fuzzy tumor boundaries, peripheral enhancement in the AP, and peripheral low signal in the HBP. We established the radiomics model by Consent for publication The data and images in this study have deleted the information of patients, protecting the privacy of patients, and can be used for publication.
Availability of data and materials The data that support the ndings of this study are available from Huiying Medical Technology, but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of Fang Wang and Yuwei Xia.
Competing interests There is no nancial or non-nancial competing interests in this study.
Funding National Scienti c and Natural Foundation of China 82060310 Authors' contributions Xialing Huang and Liling Long contributed to the methods of this study. Xinping Ye and Zili Lv contributed to the operation of the study. Ling Zhang, Muliang Jiang and Yidi Chen contributed to the data collection of this study. Fang Wang and Yuwei Xia contributed to the data analysis of the study. Xialing Huang, Jieqin Wei and Liling Long contributed to writing this manuscript.