Predicting the prognosis of hepatocellular carcinoma with the treatment of transcatheter arterial chemoembolization combined with microwave ablation using pretreatment MR imaging texture features

Objective To investigate the prognostic value of baseline magnetic resonance imaging (MRI) texture analysis of hepatocellular carcinoma (HCC) treated with transcatheter arterial chemoembolization (TACE) and microwave ablation (MWA). Methods MRI was performed on 102 patients with HCC before receiving TACE combined with MWA in this retrospective study. The best 10 texture features were screened as a feature group for each MRI sequence by MaZda software using mutual information coefficient (MI), nonlinear discriminant analysis (NDA) and other methods. The optimal feature group with the lowest misdiagnosis rate was achieved on one MRI sequence between two groups dichotomized by 3-year survival, which was used to optimize the significant texture features with the optimal cutoff values. The Cox proportional hazards model was generated for the significant texture features and clinical variables to determine the independent predictors of overall survival (OS). The predictive performance of the model was further evaluated by the area under the ROC curve (AUC). Kaplan–Meier and log-rank tests were performed for disease-free survival (DFS) and Local recurrence-free survival (LRFS). Results The optimal feature group with the lowest misdiagnosis rate of 8.82% was obtained on T2WI using MI combined with NDA feature analysis. For Cox proportional hazards regression models, the independent prognostic factors associated with OS were albumin (P = 0.047), BCLC stage (P = 0.001), Correlat(1,− 1)T2 (P = 0.01) and SumEntrp(3,0)T2 (P = 0.015), and the prediction efficiency of multivariate model is AUC = 0.876, 95%CI = 0.803–0.949. Kaplan–Meier analyses further demonstrated that BCLC (P < 0.001), Correlat(1,− 1)T2 (P = 0.023) and SumEntrp(3,0)T2 (P < 0.001) were associated with DFS, and BCLC (P = 0.007) related to LRFS. Conclusions MR imaging texture features may be used to predict the prognosis of HCC treated with TACE combined with MWA.


Introduction
Hepatocellular carcinoma (HCC) is a highly aggressive malignant tumor, and one of the leading causes of cancer death. It ranks eighth in incidence for women and fifth for men with an annual incidence of more than 660,000 new cases worldwide [1,2]. Transarterial chemoembolization (TACE) is the first treatment option for patients with unresectable HCC [3,4]. It is widely accepted as a means to control tumor growth and to prolong survival in patients with unresectable HCCs. But the complete necrosis rate of tumors after TACE is between 10 and 0% [5,6]. Failure to completely occlude the tumor-supplying artery because of angiogenesis around the residual tumor is the major cause of treatment failure. Hence, combination strategies that use both embolization and ablative methods, such as microwave ablation (MWA), have emerged to improve the clinical outcomes of TACE [7].
Previous studies on HCC have reported that the clinical efficacy of TACE combined with MWA is better than that of either therapy alone [8,9]. Assessing survival outcomes using clinicopathological data have limited prognostic value due to a lack of detailed quantitative parameters. Identifying reliable quantitative prognostic markers, therefore, remains a difficult but essential goal.
Texture analysis evaluates the heterogeneity of a tumor by quantifying the gray level intensity or position of the pixels in an image [10,11]. Recent studies showed that texture analysis could offer information on the tumor microenvironment and help predict pathological characteristics, overall survival (OS), and response to therapy [11][12][13]. Magnetic resonance imaging (MRI) is widely used to detect and characterize liver lesions, and to monitor and predict the treatment response of hepatic tumors [14,15]. However, to the best of our knowledge, no study has used MRI-based texture analyses to predict prognosis after TACE combined with MWA. The purpose of our study is to assess the value of pre-therapeutic MRI texture analysis in predicting the prognosis of HCC after combination therapy.

Patients
This retrospective single-center study was approved by the Medical Ethics Committee of our institution, and the requirement for informed consent was waived.
The study population consisted of 102 patients diagnosed with HCC according to the American Association for the Study of Liver Disease between January 1, 2013, and September 1, 2018. All patients were treated with TACE combined with MWA in our hospital (Fig. 1). All included Patients' characteristics are shown in Table 1. The inclusion criteria were as follows: (1) no previous treatment; (2) Barcelona Clinic Liver Cancer (BCLC) stage: 0, A, or B; (3) postoperative survival > 2 months; and (4) received TACE and MWA at our institution. Exclusion criteria were as follows: (1) initially diagnosed with CT, not MRI; (2) HCC after surgical treatment; (3) with history of other cancers; (4) death unrelated to HCC; (5) Lost to followed-ups; (6) with irregular follow-ups, no sufficient data for evaluating OS and prognostic factors; and (7) Serious MR image distortion.

Candidate clinical factors
We chose the following clinical features for the Cox proportional hazard models: age, sex, hepatitis B viral infection (or other serotypes); Barcelona Clinic Liver Cancer (BCLC) stage (0, A, or B); Child−Pugh class (A, B, or C); Fig. 1 Flowchart for screening HCC patients treated with TACE and MWA in our hospital maximum diameter (MD) of the lesion, number of lesions (1, or > 1); proximity to a large vessel (yes or no: yes = tumor margin is less than 5 mm from the portal vein, hepatic vein or inferior vena cava and their branches (larger than 3 mm in diameter) or no = tumor margin is more than 5 mm from the large vessels); proximity to extrahepatic organs (yes or no: yes = tumor margin is less than 5 mm from the gastrointestinal tract, liver capsule, diaphragm, and kidney or no = tumor margin is more than 5 mm extrahepatic organs); Alpha-fetoprotein level (AFP ≤ 20 ng/mL, 20-200 ng/mL or ≥ 200 ng/ mL); alanine aminotransferase (ALT ≤ 40U/L or > 40U/L); total bilirubin (TBIL ≤ 20 μmol/L or > 20 mol/L); glutamyl transferase (GGT ≤ 50 U/L or > 50 U/L); albumin (ALB ≤ 35 g/L or > 35 g/L); alkaline phosphatase (ALP ≤ 65 U/L or > 65 U/L) and prothrombin time (PT ≤ 13/s or > 13/s).

Therapy procedure
TACE was performed within 2 weeks after the diagnosis of HCC. Patients were infused with lobaplatin (50 mg/m 2 ), and then iodized oil emulsion mixed with epirubicin (30 mg/m 2 ); a microcatheter was then inserted into the tumor feeding artery. If necessary, gelatin sponge particles (150-350 μm) were injected until the flow was static. Liver and kidney functions were evaluated after TACE to ensure safe MWA. CT-guided MWA was sequentially performed at approximately 7 days after TACE. One or two 14 G antennae were inserted deep into the target lesion. The microwave power was set at 60-80 W, and the procedure lasted 10-20 min. For tumors with clear boundaries, the ablative volume enveloped the entire tumor including a 0.5-1.0 cm margin surrounding normal tissue. For tumors with irregular shapes or with obscure boundary, the ablative volume enveloped the entire tumor with a margin of 1.0 cm or more. Multiple overlapping ablations were used for tumors > 3.5 cm. For irregular tumors larger than 5.0 cm, enhanced CT within 3-7 days after the treatment was used to detect any residual viable  tissue that would require the second MWA. Vital signs such as blood pressure, heart rate, and oxygen saturation were monitored during the procedure. Hepatoprotective, antiinflammatory, analgesic, and symptomatic treatment were prescribed after MWA.

Follow-up
The patients were followed up by telephone or clinical visits 4 weeks after MWA and then every 3 months. Physical examination, hepatic function tests, AFP level, and triphasic contrast-enhanced CT or MRI were reviewed. The decision was made regarding treatment response, evidence from current guidelines, and the patients' status and intention to treat. For patients with tumor recurrence, an effective treatment plan was determined by our multidisciplinary team (MDT). Tumor recurrence included local and intrahepatic recurrence. Local tumor recurrence was defined as the presence of enhancement within or around the treated area, and intrahepatic recurrence was the presence of enhancement outside the treated area of the tumors > 1 month after treatment. OS, local recurrence-free survival (LRFS), and disease-free survival (DFS) were also assessed. All recurrences were confirmed by CT or MRI. OS was defined as the time from baseline MRI to death or end date. LRFS was defined as the time from TACE to local recurrence, death, or end date. DFS was defined as the time from TACE to local and intrahepatic recurrence, death, or end date. Patients were followed up until death or September 1, 2018, if they were still alive.

Texture features extraction and selection
All the BMP format images, including T1-weighted images (T1WI), T2-weighted images (T2WI) and contrast-enhanced T1WI in portal venous phase (PVP), were transferred into the MaZda program (http://www.elete l.p.lodz.pl/progr amy/ mazda /index .php?actio n = mazda) for texture analysis. The region of interest (ROI) on the MR parametric maps cannot be utilized automatically by MaZda. Thus, one radiologist (J. L, with 11 years of experience in MRI)-blinded to the clinical and pathological findings-manually traced the tumor border on each axial map, to obtain the corresponding two-dimensional (2D) ROI for each map. Both the most superior and the most inferior slices for each tumor were excluded to avoid volume averaging. Based on all the ROIs from the tumor, a three-dimensional (3D) volume of interest (VOI) was generated automatically (Fig. 2). For each VOI, a total of 229 texture features was extracted automatically by MaZda [17]. Nine first-order texture features were described by the histogram of the signal intensity values of Fig. 2 An example of ROI segmentation and VOI generation on T2WI. a Shows that the 2D region of interest (ROI) was delineated manually on a T2W image. b Presented that 3D view was generated automatically based on all 2D ROIs of the tumor pixels in the VOI. 220 s-order texture features (gray level co-occurrence matrix features, GLCM) were derived from 20 co-occurrence matrices produced from 4 directions and 5 inter-pixel distances in the VOI ( Table 2).
The discriminant analysis was done with the Mazda software. The best 10 texture features that predicted 3-year survival were screened out using the following statistical methods: Fisher coefficient, classification error probability with average correlation coefficients (POE + ACC), and mutual information coefficient (MI), respectively. Then texture classification was done using the B11 module in the Mazda software. Three different methods, including the principal component analysis (PCA), linear discriminant analysis (LDA), and nonlinear discriminant analysis (NDA), were applied to calculate the error rate for predicting 3-year survival (Fig. 3). The optimal feature group with the lowest misdiagnosis rate was obtained on one MRI sequence and was used for further analysis.

Statistical analysis
SPSS, version 20.0 (IBM SPSS, Armonk, NY, USA) and the R software, version 3.5.1 (R Foundation for Statistical Computing, Vienna, Austria) were used for statistical analyses. P < 0.05 was regarded as statistically significant. The differences in the values of texture features in the optimal feature group dichotomized by 3-year survival were  investigated using independent-sample t-tests. The receiver operating characteristic (ROC) was used to explore the diagnostic performance of these identified texture parameters by independent-sample t-tests, and to determine the cutoff value that would yield the best sensitivity and specificity to predict 3-year survival. Univariate analysis was performed using the Kaplan-Meier method and log-rank test concerning 16 clinical factors affecting survival. The texture features with statistical significance were divided into two groups based on the cutoff value. These significant texture features, and other significant clinical factors (screened by univariate analyses), were entered into a multivariate regression analysis by Cox proportional hazards model to predict OS (Method: forward LR; probability for stepwise: entry variables ≤ 0.05, removal variables > 0.1). To further identify the predictive performance of the multivariate Cox regression models for OS, we applied the area under the ROC curve (AUC). Survival curves for LRFS and DFS were obtained using the Kaplan-Meier method and log-rank tests.
The best 10 texture features were obtained for each MRI sequence, and the optimal feature group with the best predictive performance was obtained on T2WI. MI combined with NDA feature analysis, showed that the group with optimal features has the lowest misdiagnosis rate (8.82%) ( Table 3).

Discussion
TACE sequentially combined with MWA has been widely accepted as an effective treatment option for surgically unresectable HCC. To evaluate the prognosis of HCC after this combination treatment, we collected clinical data and pre-therapeutic MRI texture features of 102 patients. In our study, texture features obtained from T2W images had the least misdiagnosis rate of 8.82%, and therefore, had the best performance in predicting 3-year survival. This result is consistent with a previous study [18]. T2WI has a dynamic range of images with the lowest signal intensity being 0 and the highest determined by bile. Since the highest and lowest signal intensities can always be determined by the above two elements, the calculation of texture features from T2WI may be more reliable than from T1WI and enhanced T1WI.
Medical images provide sufficient information about tumor morphology as well as heterogeneity to enable tumor characterization and prognostication. Texture analysis provides crucial information regarding a tumor by evaluating its heterogeneity [19]. Texture quantification of MR images is useful in characterizing liver lesions, especially the atypical HCC in a non-cirrhotic liver where the diagnosis can be challenging even with the liver-specific contrast-enhanced MRI [20]. The clinical value of texture analysis cannot be overemphasized: it has been used to predict the HCC response after TACE [21], to determine the most appropriate treatment option between liver resection and TACE [22]and prevent unnecessary treatment.
In our research, three texture feature classifications (SumEntrp, Entropy, and Correlat) are vital prognostic factors for HCC treated with TACE and MWA. Our findings are similar to previous studies, which showed that high heterogeneity on MRI maps is usually a sign of poor outcome   [24] suggested that higher Entropy on T2WI exhibited poorer recurrence-free survival in patients with breast cancer. Also, compared with other conventional prognostic parameters, Entropy appears to be the most robust and strongest independent predictor of 2-year progression-free survival in patients with non-small cell lung cancer [25]. Meyer et al. found that Correlat had the best correlation with Ki67 index (r = 0.75), which was the clinically most relevant marker for predicting the proliferative activity of tumors [26]. Texture features are classified into first-, second-and higher-order statistics. The first-order statistics, based on histogram analysis, reflect the intensity distribution of a Note. The number in the parentheses was the patients with HCC recurrence (a, c, d) or local recurrence (b) in during the follow-up Period 1 3 VOI. The second-order statistics (an example is GLCM), such as SumEntrp, Correlat and Entropy, can reflect the spatial relationship between gray values; and is promising in predicting the therapeutic response and prognosis of many kinds of tumors [27,28]. For the first-order texture features, previous studies [29,30] showed that Variance and Skewness indicate high tumor heterogeneity, and predict poor patient prognosis. However, none of them is included in the optimal feature group in our study because GLCM provides more texture information about tumor heterogeneity [31,32] that is superior to the first-order features in predicting the prognosis of HCC. SumEntrp (3,0) and Correlat (1,− 1) were the independent predictors of OS in patients with HCC after TACE combined with MWA. Therefore, the GLCM features may be as powerful as clinical variables in predicting the prognosis of malignancies including HCC.
The BCLC stage and albumin were independent prognostic factors associated with OS, and the BCLC stage was also associated with DFS. The BCLC staging system is widely used by clinicians because it incorporates multiple variables, including tumor size, hepatic function, and performance status of patients with HCC [33]. Our analysis showed that OS of patients with BCLC stage 0 (n = 7) and A (n = 64) was longer compared to patients with BCLC stage B (n = 31). DFS and LRFS of patients with BCLC stage 0 (n = 4 (DFS); n = 0 (LRFS)) and A (n = 33 (DFS); n = 2 (LRFS)) were longer than that of stage B (n = 26 (DFS); n = 6(LRFS)). Also, the prognosis of HCC is closely related to autoimmunity and the level of inflammatory response. Preoperative ALB reflects liver function status and immune level. Some studies have suggested that the ALB level is an independent factor influencing the prognosis and recurrence of HCC [34]. Therefore, attention should be paid to understanding the preoperative ALB in predicting the prognosis of HCC.
Our study had several limitations. First, it was a retrospective study prone to potential selection bias. Second, we have a small sample size; 337 of 493 patients were excluded because they had dynamic CT rather than MRI before TACE combined with MWA. Third, we used a thick slice reconstruction in our study, however, the variation in the slice thickness of MR images doesn't significantly affect the robustness of texture features [35,36]. Further studies with larger patient populations, enhanced MRI features, and molecular biological indicators are needed to analyze the relevance between texture features and prognosis for HCC.
In conclusion, our study did not impose extra burdens on patients, as we used routinely acquired MR images. Furthermore, there are four parameters associated with OS, three associated with DFS, and one associated with LRFS. The present study showed that the 3D texture features based on MRI might be useful in predicting the prognosis of HCC before starting TACE combined with MWA.