Complementary Role of Computed Tomography Texture Analysis for Differentiation of Hepatocellular Carcinoma from Intrahepatic Peripheral Cholangiocellular Carcinoma on Single Portal-Venous Phase Computed Tomography

Methods: 61 consecutive patients (mean age, 66.7 years; range, 38.0-88.7 years; 45 males) with newly diagnosed, biopsy-proven HCC (n=33) and CCC (n=28), who all received a portal-venous phase contrast-enhanced CT (CECT) for primary staging, were evaluated in this study. Using a standardized imaging protocol, volumes of interests (VOIs) were set in the tumor tissue excluding adjacent structures. Based on pyradiomics library, a total of 35 textural features were extracted from the VOIs, including 1st, 2nd and higher order features, and then compared between HCC and CCC. Additionally, a tumor-to-liver ratio was calculated using ROIs set in the same position as for CT-textural analysis.


Introduction
Hepatocellular carcinoma (HCC) and cholangiocellular carcinoma (CCC) are the most frequent primary hepatic malignancies [1]. Differentiation between HCC and intrahepatic CCC is achieved by considering clinical, laboratory and imaging findings [2]. In the typical clinical setting (e.g., liver cirrhosis, elevated alpha-fetoprotein and typical enhancement patterns) the diagnosis of HCC can be set already non-invasively [3]. However, clinical settings for occurrence of both HCC and CCC may overlap or be clinically ambiguous, also tumor markers may be found negative [4]. In these cases of clinical dilemmas, imaging has evolved to a major diagnostic tool with some morphologic and enhancement patterns being considered suggestive for each of these two entities [5][6][7][8]. Computed tomography (CT) is mostly the first imaging technique used for classification of lesions which have been newly detected by e.g., ultrasound or suspected clinically; it is also the modality which most frequently detects incidental hepatic lesions.
As the best tissue contrast for abdominal CT-imaging is achieved in the portal-venous phase, most examinational protocols are constrained only to this phase. However, the portal-venous phase is at least capable to make the differentiation between HCC and CCC. HCC typically exhibits early enhancement (wash-in in the arterial phase) and early wash-out (in the late portal phase) in relation to liver parenchyma, whereas CCC mostly shows late post-equilibrium enhancement [9][10][11]. However, in the portal-venous phase both tumors are expected to increase their attenuation by >100% as a considerable overlap in tumor attenuation is the consequence and thus distinction between the two may be severely hampered [12].
Morphologic features like capsule formation, portal vein infiltration, radiologic evidence of cirrhosis, etc. are unable to distinguish between these tumors and therefore most of these tumors need to be differentiated histologically [12].
A novel non-invasive approach for tissue characterization has been increasingly tested and reported in the last years consisting of radiomics texture analysis [13]. Based on different textural features, differences in tumor attenuation as well as tumor heterogeneity, including those derived from the vascular network, are quantifiable and thus potentially beneficial for more accurate, profound tumor characterization. Due to the mentioned diagnostic challenges, we decided to retrospectively apply radiomics analysis to portalvenous CT image data in our own patient cohort consisting of both intrahepatic CCC and HCC that have been histologically proven to investigate potential discriminatory features.

Methods
This was a retrospective analysis of CT-image data in 61 consecutive patients (mean age, 66.7 years; range, 38.0-88.7 years; 45 males) with newly diagnosed, biopsy-proven HCC (n=33) and CCC (n=28); identified by a patient chart search at our institution between January 2014 and December 2018. Combined HCC and CCC were not included since too small of a number was available.
Patients gave informed consent and our institutional ethic board committee approved the retrospective data evaluation and registered this study under the number 536/2019BO2. All patients who underwent contrast-enhanced CT in the portal-venous enhancement phase (65-75s delay) using thin-slice image data acquisition were included. Following CT-examinational parameters were used: 120kV photon energy, 200-250 mAs tube current, a soft tissue image reconstruction kernel, and 1-3 mm slice thickness for image reconstruction. Patients received 100 mL iodine contrast agent (Ultravist 370, Bayer Vital GmbH) which was given intravenously at a rate of 2mL/s followed by a 30 mL saline chaser.

Radiomics Analysis
Radiomics analysis was performed using radiomics software (Siemens Healthineers, Forchheim, Germany) that is based on the pyradiomics package, a python package for the extraction of radiomics features from medical imaging [14]. Radiomics evaluation was applied on thin-slice portal-venous image data sets. Volumes of interest (VOIs) were drawn in a manual fashion on the transversal scans using the largest cross-section diameter avoiding partial volume averaging due to neighboring blood vessels, calcifications or visible necrotic areas Figure 1 and Figure 2. The procedure of VOI setting was performed by a junior radiologist with 3 years of experience together with a senior radiologist with 25 years of experience in abdominal and oncologic imaging; in ambiguous cases a consensus was reached. The textural features used (n= 35) belonged to 1st order (energy, total energy, entropy, minimum, calculation of tumor-to-liver attenuation ratio showing significantly higher tumor attenuation and tissue homogeneity in HCC.
Main points: a. Extract more information from a single portal-venous contrast phase.
b. Attempting to employ CT textural features to differentiate HCC and CCC.
c. Finding quantifiable parameters to assist clinicians in analysing the tumor's tissue composition.
d. May enable an educated evaluation of an otherwise non-differentiable hepatic lesion. maximum, mean, median, interquartile range, mean absolute deviation, robust mean absolute deviation, standard deviation, skewness, kurtosis, variance, uniformity as well as 2D and 3D features) as well as their customized features (squared, squared root, logarithmic, exponential, log. sigma and wavelet data); 2nd and higher order features (grey-level co-occurrence matrix, greylevel size zone matrix, grey-level run length matrix, neighboring grey-tone difference matrix and grey-level dependence matrix), the latter including sub features described by the pyradiomics library (supplementary Table 1).

Ratio Analysis
Median attenuation of the hepatic tumor in comparison to the liver was measured manually in Hounsfield units (HU) drawing a standardized region of interest (ROI) of minimum 3cm in diameter in the liver parenchyma excluding vessels as well as tumor. The ROI for the hepatic lesion was set in the same level as for the texture analysis whilst adapting its size individually.

Statistical Analysis
Statistical analysis was performed using Python 3. Receiver-operating characteristic (ROC) curves were generated by computing sensitivity and specificity at each observed cut-off.
The optimal cut-off values were derived from the point on the ROC curve with the maximum Youden index, which was calculated as For a combination of three features, we used the equation given In the study all feature combinations have been investigated.
Therefore, also the repetition of multiple features with high similarity was analyzed. In order to include as much information

Results
A highly significant difference between HCC and CCC at the significance level alpha (p<0.005) was observed for the 1 st order features Mean and Median. 1st order feature Maximum showed significant difference at the level alpha (p<0.05) whereas for glcm Idmn and glcm Idn we observed differences at the significance level alpha (p< 0.005) and (p<0.006), respectively ( having relatively low correlations ( Table 3). The first of these three feature combinations is of special interest as it contains both the attribute 1 st order Median, which shows good significance on a standalone basis (p<0.005) as well as the combination of attributes 1st order_Maximum, and ngtdm_Strength, which only had a p-value of 0.005 and p<0.5 on a standalone basis each, but a considerably smaller p-value (p<.0007) in this combination. Here, the formula is calculated analogously to the two textural feature combinations:

Ratio analysis
Median tumor-to-liver attenuation was 0.84 (95% CI of 0.74-0.90) in CCC and 0.85 (95% CI of 0.81-0.91) in HCC Figure 3. but also a slight decrease in specificity. The same was observed for the gray-level co-occurrence matrix (glcm) combined with the gray level run length matrix run entropy.
The grey-level dependency matrix dependence non-uniformity as well as the neighboring grey-tone difference matrix contrast is a measure of tissue homogeneity whereas the gray level run length matrix run entropy reflects tissue heterogeneity highlighting structural differences in particular of the vascular tumor network.
With the additional use of three feature combinations strong significances were found for some additional textural features Besides tumor morphology and their accompanying features such as lymph node enlargement, presence of ascites or portosystemic venous collaterals, the main differences in the microvascular network and tumor architecture are beneficial for diagnosis as they are responsible for differences in enhancement patterns and tissue heterogeneity. Previous reports suggested the use of quadruple phase CT for assessment of the evolution of the relative CT attenuation of HCC and CCC for sufficient differentiation [12].
These are in line with the conclusions of the Barcelona-2000 EASL Conference where three-or four-phase CT imaging protocols or dynamic contrast-enhanced MRI examinations has been recommended for non-invasive diagnosis of HCC [3].
Accordingly, a 66% increase in detection of HCC foci by using triphasic CECT scanning was reported by Oliver [15]. However, multiphase protocols presume elective liver diagnosis and are negatively afflicted by increased radiation dose. Moreover, the optimal delay time for assessment of both wash-in and washout contrast kinetics as well as of post-equilibrium delayed enhancement vary considerably among HCC and CCC patients [10,12]. Additionally, atypical enhancement patterns have been reported in patients with CCC and liver cirrhosis where tumors may exhibit a nodular arterial enhancement followed by washout suggesting HCC [16]. This has been observed especially in smaller CCCs. Conversely, progressive or stable enhancement on postcontrast scans has been observed in a notable proportion of small HCCs which can challenge radiologists [17]. Moreover, cirrhosis due to a hepatitis B or C virus infection and accompanying nodular changes in liver parenchyma are important risk factors not only for development of HCC but also for CCC development [18,19] further blurring the etiological boundaries. Besides, the attenuation of the liver parenchyma itself depends on the degree of steatosis, fibrosis or iron deposition (all of them known for their oncogenic effects) having a direct impact on the relation to tumor attenuation. The ancillary, morphologic findings of HCC/CCC overlap also considerably. Invasion of the portal vein reaches >40% in both HCC and CCC and is therefore inadequate for discrimination between these two primaries [8,12]. Cholestasis is more common in CCC but can occur also with larger HCCs. Some of these tumors are isoattenuated to liver parenchyma and some others do not objectively enhance [12]. reported 26% of CCCs in their study to have experienced no substantial late enhancement and correlated these findings with the degree of tumor differentiation and amount of intratumoral fibrosis [10].
All these particularities limit to some extent the use of morphologic features and such that are based only on visual assessment of enhancement-patterns, proving in particular problematic for one-phase (portal-venous) examinational CTprotocols which are very frequent in the primary imaging diagnosis. Therefore, we aimed to evaluate the potential of radiomics feature analysis for accurate differentiation between HCC and CCC irrespective of their imaging characteristics. To the best of our knowledge, there are no other reports dealing with the use of texture analysis for discrimination between these two tumor entities. In the last years, texture analysis has been employed successfully for HCC profiling, prognosis and response prediction to different local and systemic therapies. found that skewness on arterial-phase CT-images predicted disease-free-interval after surgical resection in patients with singular HCC whereas Defour et al. reported a similar significant association between skewness and recurrencefree survival after HCC-resection, but on portal-venous-CT-images [20,21]. Kirryu et al. using non-enhanced CT-image data found skewness and other textural features to also correlate with diseasefree-survival [22]. Pretreatment contrast-enhanced CT-texture parameters proved to be also valuable predictive biomarkers of survival in HCC patients treated with sorafenib [23]. Other reports discussed the use of radiomics data for prediction of response to local transarterial chemoembolization [24][25][26]. By comparison, we found only one report on cholangiocellular carcinoma and potential correlations between molecular profiling and imaging phenotypes using CT-image data [27], albeit accompanying lymph nodes have been investigated by textural analysis [28,29].
This study has some limitations. First, morphological and enhancement features of these tumors were not taken into consideration since in many of them additional enhancement phases (i.e., arterial or equilibrium phase) were not available. Second, histological tumor sub differentiation was also not considered as some tissue samples were analyzed by different institutions.
Third, all feature combinations have been investigated also for the repetition of multiple features with high similarity. Addressing the issue of multiple testing including as much information as possible but not having features with high similarity in the final equations, feature combinations with high correlations were ruled out. Furthermore, to overcome the low number of lesions several in sample out sample calculations were performed with different test and train samples in terms of size and observations.

Conclusion
CT textural analysis may aid differentiation of HCC from CCC in the single portal-venous enhancement phase compared to direct calculation of tumor-to-liver attenuation ratio indicating significantly higher tumor attenuation values and tissue homogeneity in HCC.