Kupffer Phase Radiomics Signature in Sonazoid-Enhanced Ultrasound is an Independent and Effective Predictor of the Pathologic Grade of Hepatocellular Carcinoma

We conduct this study to investigate the value of Kupffer phase radiomics signature of Sonazoid-enhanced ultrasound images (SEUS) for the preoperative prediction of hepatocellular carcinoma (HCC) grade. From November 2019 to October 2021, 68 pathologically confirmed HCC nodules from 54 patients were included. Quantitative radiomic features were extracted from grayscale images and arterial and Kupffer phases of SEUS of HCC lesions. Univariate logistic regression and the maximum relevance minimum redundancy (MRMR) method were applied to select radiomic features best corresponding to pathological results. Prediction radiomic signature was calculated using each of the image types. A predictive model was validated using internal leave-one-out cross validation (LOOCV). For discrimination between poorly differentiated HCC (p-HCC) and well-differentiated HCC/moderately differentiated HCC (w/m-HCC), the Kupffer phase radiomic score (KPRS) achieved an excellent area under the curve (AUC = 0.937), significantly higher than the other two radiomic signatures. KPRS was the best radiomic score based on the highest AUC (AUC = 0.878), which is prior to gray and arterial RS for differentiation between w-HCC and m/p-HCC. Univariate and multivariate analysis incorporating all radiomic signatures and serological variables showed that KPRS was the only independent predictor in both predictions of HCC lesions (p-HCC vs. w/m-HCC, log OR 15.869, P < 0.001, m/p-HCC vs. w-HCC, log OR 12.520, P < 0.05). We conclude that radiomics signature based on the Kupffer phase imaging may be useful for identifying the histological grade of HCC. The Kupffer phase radiomic signature may be an independent and effective predictor in discriminating w-HCC and p-HCC.


Introduction
Liver cancer is the sixth most common malignancy and the fourth most common cause of cancer-related death worldwide [1]. Hepatocellular carcinoma (HCC) represents primary liver cancer and is the second leading cancer-related mortality in China [2]. Clinically, the development of HCC is prone to metastasis and recurrence, which limits the prognosis as well as the quality of life [3]. Pathological grading is associated with intrahepatic recurrence [3]. High-grade HCC tumors have a higher risk of intrahepatic recurrence than low-grade tumors [4]. Accurate prediction of the HCC grade of differentiation might formulate the treatment strategy and predict the therapeutic outcome, prognosis, and risk of tumor recurrence [5,6].
Medical radiological imaging is integral to the routine clinical method of patients with HCC. Radiomics is a technology that extract the characteristics of radiological image quantitatively [7]. Conventional imaging evaluation provides few metrics without tumor heterogeneity information through the manual assessment of lesions by radiologists [8]. With the development of medical imaging data, radiomics are used to deeply excavate the biological characteristics of tumor imaging, quantitatively analyze tumor heterogeneity, and integrally evaluate tumor phenotype, which may be beyond conventional techniques. In fact, recent studies have developed that pathological grading is related to the radiomics algorithm acquired from magnetic resonance imaging (MRI) or computed tomography (CT), such as prediction of a pathological grade of gliomas and renal carcinoma [9,10].
Compared with CT and MRI, ultrasound (US) is simple, radiation-free, inexpensive, and commonly used to monitor the liver lesions [11]. Contrast-enhanced ultrasound score (CEUS) can observe the real-time vascular phase with microcirculation perfusion information. e microbubbles of a contrast agent called Sonazoid can be phagocytosed by Kupffer cells, which rarely exist in tumors, and then Sonazoid-enhanced US (SEUS) provides a special phase called the Kupffer phase or the post-vascular phase [12,13]. Previous studies have reported that the degree of the contrast defect in the Kupffer phase are related to histological grading of HCC, and certain quantifiable patterns of CEUS were associated with treatment outcomes [14][15][16].
To better interpret SEUS, we have, therefore, developed radiomics for evaluating the histological grading of HCC based on US, arterial phase, and Kupffer-phase by SEUS. Our study aims to evaluate the feasibility of US and SEUS radiomics models in terms of differentiation histologically grades of HCC to determine an initial prognosis of HCC.

Patients.
e institutional review board of our institution approved this retrospective study and waived the requirement to obtain written informed consent. Figure 1 shows the enrolment of patients.

Contrast-Enhanced Ultrasound
Imaging. All patients underwent conventional ultrasound in B-mode and SEUS by two sonographers with more than 5 years of experience in standard liver CEUS. All SEUS was performed by two sonographers using Aplio 500 (Canon, Honshu, Japan) with a convex probe (6C1, 1-6 MHz) and a linear probe (11L4, 4-11 MHz) and Aplio i800 (Canon, Honshu, Japan) with a convex probe (PVI-475BX, 1-8 MHz) and a linear probe (11L4, 4-11 MHz). e mechanical index (MI) for the acoustic output was set to 0.19-0.22 and the dynamic range was 65-70 dB according to the size of the lesion. Patients received a bolus intravenous injection of Sonazoid (perfluorobutane, GE Healthcare, Oslo, Norway) through a peripheral venous line, followed by 5 mL of normal saline flush. Immediately after the administration of Sonazoid, the hepatic arterials, portal veins, hepatic veins, and the normal liver parenchyma were uniformly enhanced during an early vascular phase image lasting 3 minutes. Approximately 10 minutes after injection, the liver was scanned again to observe the post-vascular phase image (Kupffer phase). e arterial phase and the Kupffer phase were obtained by scanning 15 to 30 seconds and 15 minutes, respectively.  https://www.itksnap.org) for manual segmentation, and a three-dimensional volume of interest (VOI) that covered the whole tumor was delineated in the images, respectively, segmented by a sonographer with over five years of experience in abdominal CEUS imaging. e procedure is shown in Figure 1 and details are introduced as follows.

Radiomics Analysis.
After integrating VOI that covered the whole tumor images, a three-dimensional radiomics feature was extracted from grayscale, arterial phase, and Kupffer phase images using the NUK software (novo ultrasound kit, GE Healthcare Shanghai, China). Specifically, shape-based (n � 9), first-order (n � 18), gray-level cooccurrence matrix (GLCM, n � 24), gray-level dependency matrix (GLDM, n � 14), gray-level run-length matrix (GLRLM, n � 16), gray level size zone matrix (GLSZM, n � 16), and neighbouring gray tone difference matrix (NGTDM, n � 5) features, according to the imaging biomarker standardization initiative from both original images and filter derived images were extracted. SMOTE up-and down-sampling was applied to create a balanced training dataset, which had been used in other radiomic studies.
Z-score normalization was applied to radiomic features. Each radiomic feature's association with the outcome (pathologically confirmed HCC differentiation status) were initially assessed using univariate logistic analysis. Radiomic features significant in univariate analysis were further selected using maximum relevance minimum redundancy (MRMR) to obtain 15 features most contributing to the outcome with least correlation. e final prediction of the outcome was obtained by a random forest classifier (RFC) trained on selected radiomic features, in the form of a radiomic score per lesion, which was calculated by linear combination of radiomic features with associated weights. e univariate-multivariate logistic model with an adjusted odds ratio (OR) was constructed using the radiomic score calculated from grayscale radiomic score (grayRS), arterial phase radiomic score (APRS), and Kupffer phase radiomic score (KPRS) images. Different HCC differentiations (low, medium, and high) were analyzed using "one versus rest" strategy (OvR).
Radiomic scores and the model's discrimination ability of pathological HCC differentiation were characterized using receiver operation characteristic (ROC) analysis; the area under the curve (AUC) was used to quantify model performance. Predictive performances including accuracy, sensitivity, and specificity of the predictors were calculated at the optimal decision boundary on the ROC curve determined by maximizing the Youden's index. Leave-one-out cross validation (LOOCV) was applied for model validation (Figure 2).

Statistical Analysis.
Descriptive statistics were presented in mean with standard deviation or median with interquartile rage depending on variables' normality. e Shapiro-Wilk test was used to asses normality. 95% Confidence intervals for model evaluation were calculated using the bootstrap method with 1000 random resamples. e DeLong test was used to compare AUC differences. e McNemar Chi-squared test was used to compare predictive performances.
e Hosmer-Lemeshow test was used to assess significance of model's deviation from perfect fit. Variables significant in univariate analysis were passed to multivariate analysis. A two-sided P value less than 0.05 was considered statistically significant.

Discussion
Radiomics is a noninvasive technology based on the quantitative extraction of signature from radiological imaging modalities [7]. In fact, investigators have shown that radiomics may be useful for predicting progression-free and overall survival for malignant diseases [17,18]. Recently, radiomics analysis based on ultrasound imaging technology has achieved some good results in the early diagnosis, prognosis, and prediction of diseases.
CEUS is widely used to observe microcirculation blood perfusion of liver cancer [19]. We used the radiomics method to evaluate the overall information related to the difference of grade that maybe contained in tumors by extracting multiphase CEUS imaging features. erefore, the aim of our study was to develop and validate CEUS radiomics models based on US, arterial phase and Kupfferphase for predicting the histological grading of HCC. Our study showed that for discrimination between p-HCC and w/m-HCC, KPRS showed an excellent AUC of 0.937, significantly higher than grayRS and APRS. Meanwhile, KPRS was the best radiomic score based on the highest AUC (AUC � 0.878), which is prior to grayRS and APRS for the differentiation between w-HCC and m/p-HCC. Wu et al. investigated MRI-based radiomics signatures for the HCC grade, and the AUC of model using radiomics signatures was 0.742 [20]. Our study showed that the prediction model using radiomics signatures based on KPRS (AUC � 0.937, 0.878) is prior to MRI, which means KPRS had advantages in predicting the HCC grade.
e easy-to-use graphic tool might provide important characteristics to stimulate clinical prediction. erefore, our study had potential application of SEUS in the diagnosis of focal liver lesions than conventional contrast medium and CT [21]. Moreover, with the number of focal liver lesions increased in HCC and other kind of tumor prior to different subtypes of hepatocellular adenoma [22].
Our study demonstrated that only KPRS demonstrated to be an independent predictor in univariate analysis and multivariate analysis in predicting the HCC grade. On the other hand, KPRS showed a better discrimination performance compared with the combination of clinical risk factors and KPRS, while, the results of were inconsistent with the previous studies. Wu et al. showed that the combination of MRI radiomics signatures with clinical factors could be useful for discriminating between high-grade and low-grade HCC, and both the AFP level and radiomics signatures were independent predictors [20]. Mohamed et al. demonstrated that serological markers, such as serum vitronectin and AFP, speculated a potential role in diagnosis and prognosis of HCC [23]. is is, probably, because of the different classification of pathological grade and the number of cases, HCC tumors were divided into low-grade and high-grade cases instead of using the International Working Party Classification or the Edmondson grade.
Limitations of this study should be acknowledged. First, the number of HCC cases was relatively limited, and HCC tumors were divided into well, moderate, and poorly differentiate-cases, while no ideal results were obtained discriminating between m-HCC and rest. Second, our study was performed in a single center, further multicenter cohorts might be necessary to evaluate the reliability, and to verify the generalizability of our findings. ird, the potential use of SEUS combined with gadoxetic acid-enhanced magnetic resonance may provide more characteristics to increase prediction [24]. In the future, multimodality ultrasound imaging-including color Doppler-flow imaging, ultrasound elastography, and vascular phase of CEUS imaging-will be combined to improve the performance of HCC classification.

Conclusions
In conclusion, radiomics signatures based on the Kupffer phase imaging may be useful for identifying the histological grade of HCC. Additionally, the Kupffer phase radiomic signature may be an independent and effective predictor in discriminating w-HCC and p-HCC.
Data Availability e data that support the findings of this study are available from the corresponding author upon reasonable request.

Ethical Approval
is study was approved by the Ethics Committee of the Beijing Hospital (2021BJYYEC-190-02).