Differentiation of Hepatocellular Carcinoma from Intrahepatic Cholangiocarcinoma through MRI Radiomics

1. Introduction

Hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC) are the most common types of primary liver cancer, with the former accounting for approximately 75%-85% [1,2,3], and their morbidity rates are increasing [4,5,6,7,8]. The treatment strategies for and prognosis of patients with HCC and ICC are very different [2,3,9,10,11,12,13,14,15,16]. Therefore, accurate preoperative discrimination between HCC and ICC is essential.

At present, the noninvasive differentiation of HCC from ICC remains challenging. The sensitivity and specificity of serum tumor markers, including alpha-fetoprotein (AFP) and carbohydrate antigen 19-9 (CA19-9), are unsatisfactory [17,18,19,20]. The presentation of HCC and ICC on dynamic contrast-enhanced computed tomography (CT) or magnetic resonance imaging (MRI) is mostly typical [21,22,23,24]. However, both HCC and ICC may occur in patients with chronic hepatitis, and imaging enhancement patterns may tend to be similar in some patients with both HCC and ICC [3,25,26,27,28,29]. In addition, the enhancement may be unremarkable or atypical in some HCC cases (especially cases of small, hypovascular, or sclerosing HCC lesions) [30,31,32]. Traditional medical imaging analysis relies heavily on the physician's subjective judgment and is thus prone to misdiagnosis [33]. Liver biopsy remains the gold standard for the final diagnosis, but this invasive procedure is refused by some patients [34]. Therefore, a preoperative, noninvasive method for distinguishing HCC from ICC is urgently needed.

The rapid development of artificial intelligence in recent years has led to it playing an important role in personalized precision medicine. Based on existing medical imaging modalities such as CT and MRI, an emerging technique known as radiomics [35] can be used to convert intrinsic pathophysiological information that is invisible to the human eye into high-dimensional quantitative image features, which can then be used to perform tumor classification via an analysis of the relationship between these features and clinical/genetic data [35,36,37]. Studies have shown that radiomics exhibits unique advantages in classifying the disease and predicting the prognosis of patients with liver cancer [35,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52]. However, there have been very few studies on the differentiation of HCC from ICC based on multisequence MRI radiomics to date. In this paper, the efficacy of a radiomic model based on preoperative fat suppression T₂-weighted imaging (FS-T₂WI) and dynamic contrast-enhanced MRI features in the arterial phase (A) and portal venous phase (P) for the differentiation of HCC from ICC was investigated.

2. Materials and Methods

2.1. Patients

The preoperative MRI and clinical data of HCC and ICC patients treated at the Affiliated Hospital of North Sichuan Medical College from April 2016 to December 2021 were retrospectively analyzed. The inclusion criteria were as follows: (1) a pathological diagnosis of HCC or ICC; (2) multisequence MRI of the upper abdomen performed within 4 weeks prior to surgery; and (3) no antitumor-related treatment prior to the MRI scan.The exclusion criteria were as follows: (1) combined hepatocellular-cholangiocarcinoma (cHCC-CC); (2) incomplete data or poor MR image quality; and (3) lesion diameter <2 cm or unclear lesion contours. The data of 206 patients with primary liver cancer (145 with HCC and 61 with ICC) were collected, and 177 (129 with HCC and 48 with ICC) met the criteria and were finally enrolled in this study. The patients were randomly divided at a 7:3 ratio into a training group (n=124, 90 with HCC and 34 with ICC) and a validation group (n=53, 39 with HCC and 14 with ICC) (Figure 1).

The following clinical data were acquired: age; sex; cirrhosis status; hepatitis B serological test results; pseudocapsule status; hemorrhagic necrosis status; extrahepatic metastasis status; portal vein tumor thrombus status; number of tumors; ascites status; intrahepatic bile duct dilatation status; maximum tumor diameter; abnormal prothrombin; AFP level; carcinoembryonic antigen (CEA) level; and CA19-9 level. The levels of tumor markers were measured within one week before surgery.

2.2. MRI Acquisition

MRI scans were performed using a Discovery 750 3.0 T superconductivity MRI scanner with a 32-channel phased-array surface coil (GE, USA). Prior to the MRI scans, all subjects fasted for 4 hours and received training in breathing exercises. Scan sequences included axial 3D liver acceleration volume acquisition (LAVA), FS-T₂WI, and axial 3D LAVA dynamic contrast-enhanced sequences (Table 1). Gd-DTPA at a dose of 15-20 mL was used as the contrast agent for dynamic contrast enhancement and injected into the dorsal vein of the hand at 2-2.5 mL/s using a high-pressure syringe.

2.3. Image Segmentation and Feature Extraction

The MR images of patients were imported in DICOM format into IBEX software (β1.0, http://bit.ly/IBEX_MDAnderson) for tumor image segmentation. The entire tumor volume was manually delineated as a region of interest (ROI) along the edge of the lesion layer by layer on axial FS-T₂W as well as A and P images if the clinical and pathological data of the patients were unknown (Figure 2). The gray-level run-length matrix (GLRLM), gray-level cooccurrence matrix (GLCM), intensity histogram and shape features were extracted and used to construct the FS-T₂WI, A, and P datasets.

2.4. Feature Selection

Sixty-one patients (42 with HCC and 19 with ICC) were randomly selected for intra- and intergroup consistency analysis. Interobserver consistency was assessed by comparing the segmentation results of two radiologists (observers 1 and 2, with 5 and 6 years of experience, respectively). Intraobserver consistency was assessed by comparing the segmentation results obtained by observer 2 more than one week after the initial results were obtained. The intraclass correlation coefficient was used to assess interobserver consistency, with a coefficient ≥0.75 considered to indicate good consistency. To eliminate discrepancies in the index dimension, all data were standardized by Z score. The dataset generated by each sequence was subjected to intra- and interobserver consistency tests. Features with an intraclass correlation coefficient <0.75 were eliminated.

From the stable features that remained, statistically significant features in differentiating HCC from ICC were selected using one-way statistical analysis (independent-samples t test or Mann‒Whitney U test, according to the characteristics of the data distribution) (p<0.05). To avoid overfitting, least absolute shrinkage and selection operator (LASSO) regression analysis was performed to select the core radiomic features for differentiating HCC from ICC. The regularization parameter (λ) of the selected features was adjusted by 10-fold cross-validation using the 1-standard error (1-SE) method.

2.5. Model Establishment and Evaluation

The optimal radiomic features selected from each sequence were used to establish the radiomic models (FS-T₂WI, A, and P models) by logistic regression. By integrating the optimal features, a joint radiomic model (M) was established [41]. The efficacy of the models was assessed by the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy and F1 score as determined from the logistic regression confusion matrix.

2.6. Statistical Methods

R software (4.1.2, https://www.r-project.org/) was used for statistical data processing. Specifically, the software packages "psych", "glmnet", and "pROC" were used to assess the intergroup consistency of radiomic features, to perform LASSO regression analysis and to plot ROC curves, respectively. The normality and homogeneity of variance of the quantitative data were tested by the Shapiro‒Wilk test and Bartlett test, respectively. The independent-samples t test was performed for quantitative data with a normal distribution and homogeneous variance; otherwise, the Mann‒Whitney U test was performed. Quantitative data are described by the means or medians. Categorical variables were analyzed by the chi-square test and are described by percentages. Two-sided p values <0.05 was considered statistically significant.

3. Results

3.1. Patient Characteristics

Among the 177 patients, 129 had HCC (112 men and 17 women), and 48 had ICC (19 men and 29 women). Cirrhosis occurred in 121 patients (107 with HCC and 14 with ICC), and multinodular liver cancer occurred in 65 patients (47 with HCC and 18 with ICC). The maximum tumor diameter was 6.57±3.22 cm (Table 2).

3.2. Feature Extraction and Selection

A total of 352 features were extracted from each of the FS-T₂WI, A and P datasets. Features with intra- and intergroup intraclass correlation coefficients <0.75 were eliminated, and the remaining features were further analyzed (Figure 3).

There were 327, 331 and 319 significantly different features in the FS-T₂WI, A and P datasets, respectively (p<0.05), according to the independent-samples t test or Mann‒Whitney U test. Finally, 1, 6 and 4 optimal features of these datasets were selected by LASSO regression, respectively (Figure 4 and Table 3).

3.3. Model Establishment and Evaluation

The above features selected from the FS-T₂WI, A and P datasets were used to establish each radiomic model, and then the joint model (M model) was established by integrating the 11 features into a single model. The AUC of the FS-T₂WI, A, P and M models was 0.693, 0.863, 0.818 and 0.914 in the training group and 0.690, 0.784, 0.727 and 0.802 in the validation group, respectively (Table 4 and Figure 5).

4. Discussion

MRI is characterized by high soft-tissue contrast, multiparametric and multidirectional imaging and a lack of radiation, making it the preferred imaging method for the identification and diagnosis of liver nodules [53,54]. Dynamic contrast-enhanced MRI (DCE-MRI) is superior to dynamic contrast-enhanced CT in the detection and diagnosis of small HCC lesions (maximum diameter ≤2.0 cm) [55,56]. Typical HCC displays significant heterogeneous enhancement in the arterial phase on DCE-MRI and reduced enhancement in the portal and/or parenchymal phase that is lower than that of normal liver parenchyma, resulting in a "fast-in and fast-out" enhancement pattern [21,24]. In contrast, ICC shows less obvious enhancement or heterogeneous mild enhancement in the arterial phase on DCE-MRI that gradually increases with time [22,23]. However, it is still difficult to differentiate HCC from ICC in clinical practice. Studies [26,30,31,32] have shown that small ICC lesions (diameter<3 cm) and some ICC lesions in the setting of cirrhosis (approximately 7%) show the same enhancement pattern as typical HCC lesions, and approximately 10%-20% of HCC lesions (especially small, hypovascular, or sclerosing HCC lesions) show less obvious enhancement on imaging.

Choi et al. [57] conducted gadoxetic acid-enhanced MR and dynamic CT scans to identify HCC and ICC. The results showed that portal venous phase (PVP) washout instead of conventional washout in gadoxetic acid-enhanced MRI can prevent misidentification of HCC as ICC in patients with cirrhosis; however, it reduces the sensitivity of the method for identifying HCC. Diffusion-weighted imaging (DWI) reflects the diffusion of water molecules in tissues by measuring the apparent diffusion coefficient (ADC). Wei et al. [58] and Lewis et al. [59] found that the ADC can help differentiate HCC from ICC. However, ICC has multiple cellular origins and shares similar biological behaviors to some extent with HCC; thus, the ADC of ICC can partially overlap with that of HCC. Additionally, DWI does not display small lesions well because of the limited spatial resolution, and conventional DWI is based on a monoexponential model that cannot differentiate between water molecule diffusion and blood perfusion [60]. Intravoxel incoherent motion-DWI (IVIM-DWI) can simultaneously quantify the diffusion of water molecules and microcirculatory perfusion in living tissues. A previous study [61] showed that both the ADC and Dslow values were significantly lower in HCC than in ICC, but the Dfast value was significantly higher in HCC than in ICC; furthermore, Dfast was more efficient in the differential diagnosis of HCC and ICC, and there was no significant difference in the f value between Dfast and Dslow. The value of IVIM-DWI in identifying HCC and ICC has also been reported by other scholars [58,62,63]. However, the conclusion regarding D_fast and f in distinguishing HCC from ICC remains inconsistent or controversial; thus, further research is needed. As an effective tumor imaging tool, positron emission tomography (PET)-MRI can play a role in patient management. Çelebi et al. [64] argued that PET-MRI using ¹⁸F-fluorodeoxyglucos (¹⁸F-FDG) as a tracer agent can help differentiate between HCC and ICC. However, there is a need to deeply explore whether there are great differences in FDG uptake between HCC and ICC, the accuracy of identification in some challenging cases (e.g., specific subgroups of patients in which the standard uptake value (SUV) is not a determining factor), and the optimal imaging sequence and model.

To date, few studies have investigated the differentiation of HCC from ICC based on MRI radiomics [38,40,59,65]. Liu et al. [40] adopted machine learning-based CT and MR image features in the identification of cHCC-CC, ICC and HCC. The results showed that MRI features had the highest efficacy in differentiating between cHCC-CC and non-cHCC-CC, while CT features were less valuable. Moreover, precontrast- and portal-phase CT features were superior to enhanced MRI features in differentiating between HCC and non-HCC (AUC=0.79-0.81 for MRI, 0.81 for precontrast-phase CT and 0.71 for portal-phase CT). Wang et al. [65] used MRI radiomics to preoperatively identify cHCC-CC, HCC and ICC and found that the performance of the higher-order feature-based model was better than that of the lower-order feature-based model by approximately 10% and that the former performed better in identifying HCC in the delayed phase. Lewis et al. [59] extracted first-order radiomic features from ADC data and evaluated the ability of these features and the Liver Imaging Reporting and Data System (LI-RADS) classification to differentiate HCC, ICC and cHCC-CC. The results revealed that the AUC of the combination of sex, LI-RADS grade and the fifth percentile of the ADC in the diagnosis of HCC was 0.90 and 0.89 for two independent observers, respectively,. T₂^∗WI can reflect the magnetic susceptibility variation in tissues and thus be used to assess the biological properties of tumor tissues [66]. Huang et al. [38] extracted radiomic features from T₂^∗W images and then established radiomic nomogram models combined with clinical risk factors to distinguish between HCC and ICC. The results showed that the AUC of the radiomic model was 0.90 and 0.91 in the training and validation groups, respectively; the AUC of clinical features was 0.88 in the training group and 0.83 in the validation group, and the AUC of the radiomic nomogram was 0.97 and 0.95 in the training and validation groups, respectively. Similar results were obtained by Zhou et al. [39]. However, the efficacy of a joint model incorporating multiple sequence features was not investigated in these studies.

Different kinds of information related to tumor structure can be revealed by different sequences: T₂WI exhibits the underlying tumor morphology and heterogeneity, and enhancement scans can reflect differences in the tumor blood supply. In this work, enhancements in the arterial and venous phases were combined based on T₂WI to explore the efficacy of a joint model according to the blood supply status and enhancement patterns of HCC and ICC. The results showed that while each model had the potential to identify HCC and ICC in both the training and validation groups, the joint model incorporating multiple sequence features showed the highest efficacy [40,65]. The AUC of the T2WI model was relatively low in this study, consistent with the findings of Liu et al. [67]. Therefore, the value of FS-T₂WI-based radiomics in distinguishing between HCC and ICC remains to be properly determined pending further research.

The radiomic features selected in this study were mainly GLCM and GLRLM features, textural features used to quantify tumor heterogeneity by reflecting the relationship between adjacent voxels/pixels [68], which is consistent with the results of related studies [38,40,69,70,71,72,73,74]. Histogram features show the global distribution of grayscale values in the image and can also be used to assess tumor heterogeneity [75]; Lewis et al. [59] found that the 5^th/10^th/95^th percentiles of the ADC could significantly differentiate HCC from ICC and cHCC-CC. Shape features reflect the geometric characteristics of tumors [68]; Zhao et al. [76] confirmed that HCC tends to be more spherical than ICC in terms of morphology.

This study had the following limitations. (1) In this retrospective study, many HCC and ICC patients who did not undergo preoperative MRI scans were excluded, so there may be a potential selection bias. (2) The sample was small and from a single center, and cHCC-CC and ICC types other than the mass-forming type were not included in the study. In the future, the sample size should be expanded to multiple centers for further model validation. (3) Other relevant MRI sequences were not analyzed, so their potential contributions might have been ignored.

5. Conclusions

Multisequence MRI radiomic models can be used to preoperatively distinguish between HCC and ICC, and the efficacy of these models can be further enhanced by including information from multiple sequences.