Computed tomography texture analysis to facilitate therapeutic decision making in hepatocellular carcinoma

This study explored the potential of computed tomography (CT) textural feature analysis for the stratification of single large hepatocellular carcinomas (HCCs) > 5 cm, and the subsequent determination of patient suitability for liver resection (LR) or transcatheter arterial chemoembolization (TACE). Wavelet decomposition was performed on portal-phase CT images with three bandwidth responses (filter 0, 1.0, and 1.5). Nine textural features of each filter were extracted from regions of interest. Wavelet-2-H (filter 1.0) in LR and wavelet-2-V (filter 0 and 1.0) in TACE were related to survival. Subsequently, LR and TACE patients were divided based on the wavelet-2-H and wavelet-2-V median at filter 1.0 into two subgroups (+ or −). LR+ patients showed the best survival, followed by LR-, TACE+, and TACE-. We estimated that LR+ patients treated using TACE would exhibit a survival similar to TACE- patients and worse than TACE+ patients, with a severe compromise in overall survival. LR was recommended for TACE- patients, whereas TACE was preferred for LR- and TACE+ patients. Independent of tumor size, CT textural features showed positive and negative correlations with survival after LR and TACE, respectively. Although further validation is needed, texture analysis demonstrated the feasibility of using HCC patient stratification for determining the suitability of LR vs. TACE.


INtrODUctION
Identification and quantification of tumor heterogeneity by computed tomography (CT) textural analysis shows promise in enhancing prognostic accuracy and facilitating therapeutic decision making. Such advances are particularly important for diseases such as liver cancer, which is the second and sixth most frequent cause of cancer related-death in men and women, respectively, with hepatocellular carcinoma (HCC) most common [1]. According to the Barcelona Clinic Liver Cancer (BCLC) staging system, the diameter of a single HCC may not be a contraindication for liver resection (LR) [2][3][4][5]. However, most Asian patients with HCC have diseased liver parenchyma, such as hepatitis B virus infection and/or hepatitis B virus-related cirrhosis, and LR in this population is therefore associated with a high risk of complications [6]. This consideration may alter therapy-based decision-making in cases of single HCCs > 5 cm, particularly for potential LR candidates [7,8]. Asymptomatic patients with a solitary HCC without vascular invasion or extrahepatic spread and with wellpreserved liver function could also be considered for transcatheter arterial chemoembolization (TACE) [3,5,9]. Thus, to assess whether a patient scheduled to undergo LR would be better suited for TACE and vice versa, reliable prognostic markers for patient stratification are needed.
Proposals for a subclassification system for BCLC stage B tumors have emerged in recent years. One study proposed a stratification system aimed toward tailoring therapeutic interventions based on both the evidence available to date and expert opinions [10]. Another study suggested taking the Child-Pugh score and liver transplantation status into account [11]. In clinical practice, the decision to treat with LR or TACE is made using a combination of clinical symptoms, laboratory test results, and pathological biomarkers, whereas the CT images routinely acquired during treatment and followup are largely overlooked. Conventional assessment of tumor size and enhancement of cross-sectional images are far from satisfactory in the determination of an appropriate therapeutic strategy, due to insufficient imaging of the inherent properties of the tumor and interobserver variability in image interpretation.
Radiomics is an emerging research field that aims to utilize the full potential of medical imaging [12]. This includes texture analysis, which is assumed to reflect tissue heterogeneity [13][14][15][16][17]. Heterogeneity is a well-recognized feature reflecting alterations in tissue patterns, likely occurring due to cell infiltration, abnormal angiogenesis, microvasculature and necrosis [18][19][20]. One study suggested an association between image traits, including textural features, and underlying gene expressions in HCC [21]. Another stated that the radiomic signature could be transferred from lung to head-and-neck cancer, suggesting that this signature identifies a general prognostic tumor phenotype [12]. In fact, texture analysis has shown feasibility in the differential diagnosis of liver cancers [22], hepatic fibrosis detection/staging [23,24], and prediction of postoperative hepatic insufficiency [25].
In this study, we explored texture analysis as a prognostic and patient stratification approach in the determination of the appropriate therapeutic option, LR or TACE, for patients with single large HCCs. Herein, two questions were raised: (1) are the textural parameters of the primary tumor, calculated from baseline CT, related to prognosis? (2) Does texture analysis have the potential to provide an additional view for treatment modification between LR and TACE? rEsULts Patients A total of 130 patients (86 and 44 treated by LR and TACE, respectively) were retrospectively included for texture analysis. Of these, 106 (81.5%) patients had disease progression, and 96 (73.8%) patients died by the study end date. There were no significant differences in the patient baseline demographics and characteristics (Table 1). All texture features, calculated from two sets of regions of interest (ROIs), showed excellent agreement (ICC value, 0.799-0.999).

cox regression and Kaplan-Meier analysis for Lr and tAcE
For candidate clinical and imaging variables, univariate analysis showed that corona (P = 0.057) was the only variable with a P value < 0.10 in LR, whereas in TACE, none of the variables showed significant differences ( Table 2). For textural features, nine and 21 features in the LR and TACE groups, respectively, were identified as statistically significant (Table S1).
Separated by the above-identified four textural parameters in LR and TACE, OS differed significantly for each feature whereas time to progression (TTP) did not ( Figure 1 & Table 4). There were no significant differences in patient demographics and characteristics.
In all patients, for OS, univariate Cox regression showed that BCLC, corona, and subgrouping had P-values < 0.10, and the multivariate Cox regression models confirmed that subgrouping was the only factor that was significantly associated with OS (P = 0.012). For TTP, univariate Cox regression showed that the presence of a capsule, corona and subgrouping had P-values < 0.10, and the multivariate Cox regression models confirmed that the capsule was the only factor that was significantly associated with the TTP (P = 0.021).
These results indicate that LR+ was associated with the best survival, followed by LR-and TACE+ (P = 0.920 and 0.854 for OS and TTP, respectively, in LRvs. TACE+), whereas TACE-was associated with the worst survival. Thus, the feasibility of texture features in patient stratification and determination of the most suitable therapy (LR or TACE) was partly confirmed; however, further validation was still considered necessary.

Further validation
Since wavelet-2-V (filter 1.0) and wavelet-2-H (filter 1.0) were not normally distributed among the subgroups, Kruskal-Wallis H was used for further analysis.
First, wavelet-2-V (filter 1.0) was compared between LR+ and TACE+, as well as between LR+ and TACE-. The results showed that the value of LR+ was   Figure 3A). Second, wavelet-2-V (filter 1.0) was compared between LR-and TACE+, as well as between LR-and TACE-. The results showed that the value of LR-was similar to that of TACE+ (median, 13.1050 vs. 12.8860, P > 0.999), but lower than that of TACE-(median, 13.1050 vs. 18.3490, P < 0.001). Therefore, if LR-patients are treated by TACE, their survival would be similar to that of TACE+ patients and better than that of TACE-patients, without compromise of OS ( Figure 3B).
Third, we compared wavelet-2-H (filter 1.0) between TACE-and LR-, as well as between TACEand LR+. We observed that TACE-and LR-showed a significant difference (median: 15.3530 vs. 11.8430, P < 0.001), whereas TACE-and LR+ did not (median: 15.3530 vs. 15.7260, P > 0.999). Therefore, if TACE-patients are treated by LR, their survival would be similar to that of LR+ patients and better than that of LR-patients, and their OS would be considerably improved ( Figure 3C).
Lastly, we compared wavelet-2-H (filter 1.0) between TACE+ and LR-, as well as between TACE+ and LR+. We observed that TACE+ and LR+ showed a significant difference (median: 11.5270 vs. 15.7260, P < 0.001), whereas TACE+ and LR-did not (median: 11.5270 vs. 11.8430, P > 0.999). Therefore, if TACE+ patients  are treated by LR, their OS would be similar to that of LR-patients and worse than that of LR+ patients, with no extension of survival ( Figure 3D).
Accordingly, when the TACE group was separated by another prognostic indicator, wavelet-2-V (filter 0), similar results were consistently noted ( Figure 2).

DIscUssION
In this study, we took six typical HCC subject image features and textural features into account to determine whether or not they could be used to assist in therapeutic decision-making and optimization. Corona and 29 textural parameters (nine in LR and 20 in TACE) had P -values < 0.10 in the univariate Cox regressions for OS. Sequentially, multivariate Cox regressions and Kaplan-Meier analyses identified four parameters (one in the LR group and three in the TACE group) related to OS. Filter 1.0 was the best filter, as it showed significant results in both the LR and TACE groups, which was consistent with the findings of published studies [15,16]. The reason for this result might be that textural features at filter 0 tend to reflect radiologists' impressions of image quality, which could be influenced by image noise. By using filters at larger scales (filter 1.0, 5 pixels), subjective bias might be alleviated, and underlying biologic heterogeneity could be enhanced [14].
In the subgroup comparisons, we noted that TACE would have severely compromised OS in LR+ patients,    with lower values might be recommended for TACE ( Figure 3). Additionally, similar conclusions could also be drawn if separated by the receiver operating characteristic curve threshold ( Figure S1 & S2).
Although BCLC stage C used to be considered a contradiction for both TACE and LR in HCC, recent studies showed that both TACE and LR could provide survival benefit [26][27][28]. Nevertheless, patient selection was crucial before LR or TACE. So, we included HCCs in BCLC stage C in our study. Furthermore, in clinical observations, sorafenib alone seldom led to necrosis in intrahepatic/extrahepatic lesions or shrinkage in thrombosis. Given that no more effective treatments are recommended by the BCLC staging system, BCLC stage C HCC patients would have significantly shorter survival times than stage B patients. However, in this study, patients in BCLC C stage only had thrombosis at bifurcations. In this situation, the branches involved could be removed by resection or embolized by TACE. Therefore, intrahepatic lesions and branch vascular invasion could be treated simultaneously. This might be one explanation for why the BCLC did not have prognostic value in this study. Nevertheless, our results should be considered as preliminary, and further study is warranted.
In HCC prognosis, tumor stage and accurate evaluation of the liver-function reserve need to be incorporated [5]. Thus, the presence of cancer-related symptoms, liver function, alpha-fetoprotein, and Child-Pugh score were adjusted for, with no significant differences observed, and subgrouping according to texture analysis proved to be the only factor significantly related to OS. For TTP, subgrouping showed significant differences in the Kaplan-Meier survival curves, but not in the Cox regression. The reason for this observation might be that in the initial screen by Cox regressions, we used survival status (OS) as the event, which probably excluded some parameters related to TTP.
Texture analysis is associated with challenges in image acquisition. In a previous phantom study, texture parameters were demonstrated to be relatively sensitive to tube voltage, but to be independent of the tube current [29]. Additionally, one study showed that hepatic texture features were less sensitive to changes in CT acquisition parameters [30]. Slice thickness is another major determinant of textural parameters [31], with one study revealing that a slice thickness of ≤ 3 mm was optimal for feature grading [32]. Thus, we carefully excluded images outside this criterion in the present study, which might have partly reduced the influence of textural parameter reproducibility in prognostic evaluation.
This retrospective study had some limitations. First, this study included a relatively small sample size. However, in an attempt to control for possible confounding effects, patients with multiple lesions or extrahepatic metastasis were excluded. Second, the retrospective design of this study did not include some potential confounding factors. In particular, the prevalence of comorbidities that might influence liver texture, such as diabetes, alcoholic liver disease, and early cirrhosis was unknown and needs to be assessed in future studies. Further, the possibility of selection bias could not be eliminated. Finally, in this study, all ROIs were manually drawn by the two radiologists rather than by automatic segmentation. However, excellent inter-observer agreement was observed. Future adoption of a more robust algorithm is warranted.
In conclusion, textural variations on baseline CT images might offer more thorough insight for HCC prognosis. Additionally, detailed grouping by wavelet features showed the feasibility of this method in patient stratification. Though further validation is still warranted, texture analysis could potentially be used to inform the LR vs. TACE decision-making strategy.

Patients
This study was approved by the Ethics Committee of Guangdong General Hospital. Informed consent was waived due to the retrospective design of the study, and all patient records and information were anonymized and de-identified prior to analysis.
Between September 2009 and December 2014, 130 patients with a single large HCC ( > 5cm) initially treated by LR or TACE were enrolled (Figure 4). The time interval between baseline CT and initial treatment was less than 14 days. For BCLC stage C, only patients with branch vascular invasion were included, whereas those with extrahepatic metastasis and main portal vein thrombosis were excluded due to limited efficacy of LR/TACE for these patients. For all enrolled patients (if still alive by the study end date of March 2015), at least a three-year follow-up was required; patients diagnosed after March 2012 without death were excluded from this requirement.
In this study, we employed the BCLC classification instead of TNM, as the BCLC classification is more informative than TNM regarding survival outcomes and therapeutic strategies [2,3]. In the BCLC classification, disputes exist on the classification of single HCCs > 5 cm without vascular invasion and extrahepatic metastasis as stage A or B [5,28]. Herein, we used stage AB, as proposed by a previous study [28]. As a result, single HCCs > 5 cm with well-preserved liver function (Child Pugh A-B cirrhosis, and PST < 1) were classified as stage AB (without vascular invasion) and stage C (with vascular invasion). Patients with PST = 1 but without vascular invasion were still classified as stage AB, which was in accordance with two studies [28,33].
Anatomical or non-anatomical LR was performed with a margin > 10 mm. All TACE procedures in this study were performed by Seldinger's technique, with epirubicin, lobaplatin and lipiodol mixed as the embolic agent.

Follow-up and endpoint
The follow-up interval was 4-8 weeks, and included routine laboratory tests, chest X-ray and abdominal CT. Additional CT or magnetic resonance imaging was routinely performed if extrahepatic metastasis was suspected.
The primary endpoint was OS, and the secondary endpoint was time to progression (TTP). Disease progression for TACE was defined as an increase of at least 20% in the diameter of a viable target lesion according to the modified Response Evaluation Criteria in Solid Tumors. Disease progression for the LR group was defined as intrahepatic or extrahepatic recurrence.

ct examination
All baseline images were derived from our picture archiving and communication system. Portal-phase CT of liver was obtained by the same scanner (LightSpeed VCT 64; GE Medical Systems, Waukesha, WI). After administering iopamidol (370 mg of iodine/mL, Iopamiro; Bracco, Milan, Italy), a non-ionic contrast medium, at 1.5 mL/kg (maximum dose, 100 mL) with a double-tube highpressure syringe at 3.5 mL/s, hepatic imaging acquisition was performed at fixed time points at the portal venous phase at a 70 sec delay. The scan parameters were as follows: 120 kV; automatic mA, 80-500 mA; noise index: 7; pitch/table speed = 0.984/39.37 mm/rot; rotation time, 0.6 s; field of view, 300-450 mm; matrix, 512 mm. A slice thickness of 1.25 mm was routinely reconstructed with soft kernels.

texture analysis methodology
For each pre-treatment examination, 1.25-mm axial images obtained at the portal venous phase through the largest cross-sectional area of the tumor were selected and transferred to two personal computers for texture analysis. The process of texture analysis comprised three steps: (1) image filtration, (2) wavelet analysis and (3) feature extraction. The first two steps were performed using MATLAB 2014a software (MathWorks Inc., Natick, MA).
(1) Image filtration: Laplacian of Gaussian (LoG) spatial band-pass filters were used to reduce the sensitivity to noise. Filter width and sigma (σ) are the two parameters that characterize LoG filter weighting. Three σ values (0, 1.0, and 1.5) and a single filter-width of σ*5 pixels were used (Table S2). Pixels with attenuation of less than -50 HU were removed. The filtration process produced a series of images displaying textual features at different filters.
(2) Wavelet analysis: The use of the wavelet transform for texture analysis was first proposed by Mallat [35]. This transform provides a robust methodology for texture analysis in different scales. Initially, it decomposes each image and receives its texture by using a series of elemental functions called wavelet and scaling, where "s" governs the scaling and "u" the translation, as follows: As a result, the Haar wavelet transform decomposes each original image into nine images with different scales, called trends and fluctuations: The former are averaged versions of the original image, and the latter contains the high frequencies.
Each image is decomposed into 1, 2, or 3 levels and reconstructed in three directions (diagonal, horizontal and vertical).
(3) Texture feature extraction: Two radiologists (Readers 1 and 2, with 5 and 4 years of experience in abdominal CT interpretation, respectively) independently performed textural feature extraction and quantification using ImageJ software (National Institutes of Health, Bethesda, MD). For each reader, a user-defined irregular ROI was drawn manually around the largest crosssectional tumor outline and copied to the nine derived texture feature maps. Subsequently, the values of the texture features were measured and saved for further analysis.

statistical analyses
The Shapiro-Wilk test was applied to assess normality. Differences in patient demographics and characteristics for those undergoing LR or TACE were tested using independent-sample t-tests, Mann-Whitney U tests and Chi-square tests. Inter-observer agreement on textual features was evaluated using intraclass correlation coefficients (ICCs) [36].
Patient demographics and subjective imaging features were included for adjustment in the analyses. Univariate Cox regression was used as a preliminary screening of candidate variables. Variables of statistical significance in the univariate analysis (P < 0.10) were used as input variables for the subsequent multivariate www.impactjournals.com/oncotarget Cox regression models (Forward: LR method). Textural features at each filter were tested in separate models to assess the independent effects of the CT texture of the primary tumor on OS. Thus, six multivariate models were created (one per group per filter; LR and TACE groups, and three filters). Afterwards, the median values of the independent texture parameters were used to separate patients in the LR and TACE groups for subsequent Kaplan-Meier analysis.
To explore the potential role of texture features in deciding between LR and TACE, patients were first divided into two subgroups according to the identified prognostic markers in the LR and TACE groups. Cox regression for all patients was performed to determine whether subgrouping was an independent factor for OS and TTP. Next, one way-ANOVA or Kruskal-Wallis H was used to compare the identified textural parameters among the subgroups. Post hoc multiple comparisons were performed using Bonferroni's correction or Dunnett's T3 test.
The thresholds of the identified factors in Cox regression models were also determined using standard receiver operating characteristic curves. Figure S1 & S2 contain detailed discussions regarding this approach.
All statistical analyses were performed with SPSS 20.0 (IBM SPSS Statistics, Armonk, NY). A two-tailed P value of less than 0.05 was considered statistically significant.