Tumor heterogeneity of pancreas head cancer assessed by CT texture analysis: association with survival outcomes after curative resection

The value of image based texture features as a powerful method to predict prognosis and assist clinical management in cancer patients has been established recently. However, texture analysis using histograms and grey-level co-occurrence matrix in pancreas cancer patients has rarely been reported. We aimed to analyze the association of survival outcomes with texture features in pancreas head cancer patients. Eighty-eight pancreas head cancer patients who underwent preoperative CT images followed by curative resection were included. Texture features using different filter values were obtained. The texture features of average, contrast, correlation, and standard deviation with no filter, and fine to medium filter values as well as the presence of nodal metastasis were significantly different between the recurred (n = 70, 79.5%) and non-recurred group (n = 18, 20.5%). In the multivariate Cox regression analysis, lower standard deviation and contrast and higher correlation with lower average value representing homogenous texture were significantly associated with poorer DFS (disease free survival), along with the presence of lymph node metastasis. Texture parameters from routinely performed pre-operative CT images could be used as an independent imaging tool for predicting the prognosis in pancreas head cancer patients who underwent curative resection.

Correlation of recurrence with texture features. Regarding the CT texture features without filtration and with the various filter values, the areas under the curve (AUCs) and the optimal cut-off values for diagnosing recurrence determined by ROC curve analysis are summarized in Table 2. Without filtration and with fine (1.0) and medium (1.5 and 2) filter values, the recurrence group showed significantly lower averages, contrast and standard deviations, and higher correlations than the non-recurrence group. Only the average and contrast were significantly different between the two groups with the coarse filter (2.5) value. Applying cross-validation using Leave-one-out cross validation (LOOCV) model yielded the optimal cut-off values in concordance with the previous results (Supplementary Table S1).  Tables 3 and 4, respectively. The univariate Kaplan-Meier analysis with the log-rank test for DFS showed significant differences for the presence of lymph node metastasis, the dichotomized average, contrast, correlation, and standard deviation with no filter and fine to medium filters and for the dichotomized average with the coarse filter (Table 3, Fig. 1). The cross-validation of Kaplan-Meier analysis by LOOVC model for DFS results were in line with the previous results ( Table 3, Fig. 2). It showed statistically significant difference for most of the features except for correlation in the filter value of 0, average in the filter value of 1, standard deviation in the filter value of 1.5.
In the multivariate Cox proportional hazards model, the presence of lymph node metastasis was an independent factor that showed a significant association with DFS regardless of the applied filter (hazard ratio [HR], 1.957 to 2.181). Various texture features-including average filter values of 0, 1, and 2.5; standard deviation in the filter values of 0 and 2; contrast in the 1.5 filter value; and correlation in the filter value of 1-served as independent prognostic factors for predicting poorer DFS (Table 4). Overall, homogeneous texture features (lower standard deviation and contrast and higher correlation) with a lower average value from the texture analyses were significantly associated with poorer DFS.

Discussion
In this study, we have demonstrated the prognostic value of texture features of preoperative CT images using histograms and GLCM analyses in patients with pancreatic head cancer who have undergone curative resection. Our results show that lower average values with homogeneous features (lower standard deviation and contrast and In our study population, lower average values (without filtration and with filters 1.0 and 2.5) of pancreatic head cancer were found to be negative prognostic factors for DFS after curative resection in a multivariate Cox proportional hazards model. As the pixel histogram average represents the brightness or mean gray-level intensity of a region, a lower average on the texture analysis indicated a lesion with low attenuation. One of characteristic pathologic features of pancreas cancer is the presence of intense fibrosis in the tumor, which is known as desmoplastic reaction 21 . Previous studies have demonstrated that scirrhous carcinomas with abundant fibrosis and relative sparse tumor cells in the stomach, bile duct, breast, and colon have a poor prognosis 22 evidence is accumulating that the fibrous component of a tumor correlates with its malignant behavior and contribute to therapeutic resistance 26 . Although the significance of the fibrotic component of pancreas cancer is still unclear, peritumoral fibroblasts in pancreatic cancer have been shown to overexpress SPARC (secreted protein acidic and rich in cysteine), which is a marker of poor prognosis when expressed in the stroma 27 . Furthermore, a desmoplastic reaction in pancreas cancer is thought to be responsible for metastasis, as well as chemotherapy resistance, by reducing the amount of drug delivered to the tumor 28 . Because pancreatic cancer frequently has an abundant fibrotic stroma, which is seen as a hypo-attenuating mass in the early arterial phase with progressive delayed enhancement 29,30 , we believe that the lower average observed on the pancreas phase images reflects pancreatic cancer with abundant desmoplastic reactions. Other studies have suggested that iso-attenuating pancreatic cancers on early-phase images tend to display less desmoplastic change within the mass and show better survival outcome 25,[31][32][33][34][35][36] . Studies involving diffusion MRI have concluded that the degree of fibrosis in pancreatic cancer cases correlates with diffusion restrictions related to poor prognosis 37,38 and that it could be used to monitor treatment response 39 . We speculate that pancreatic head cancer with a lower average may reflect an imaging phenotype   40 , who demonstrated that hypo-attenuating pancreatic cancer in the portal-venous phase on CT scans showed shorter DFS. However, the contrast between normal parenchyma and pancreatic cancer is greater in the pancreatic phase than in the portal-venous phase, and tumors normally demonstrate peripheral enhancement of the tumor in the portal-venous phase [41][42][43] . Therefore, our data obtained from the pancreatic phase would be better for representing the entire tumor mass as well as the internal heterogeneity compared to the data from the portal-venous phase.
Also, interestingly, our study has revealed that both first-(a lower standard deviation without filtration and with the 2.0 filter) and second-order statistics (a lower contrast with the 1.5 filter and a higher correlation with the 1 filter) representing intratumoral homogeneity are related to poorer DFS in the multivariate Cox proportional hazards model. The first-order statistics, calculated from a histogram of pixel values, were based on the gray-level frequency distribution and represent a single pixel value rather than its spatial relation to adjacent pixels 13,44 . Instead, secondary parameters, calculated using GLCM, show the spatial relationship between one pixel and another. These secondary parameters have the advantage of being able to quantify the overall texture content 13,45 . Our study differs from the work of Cassinotto et al. 40 , who only used first-order statistics to perform a texture analysis in pancreatic cancer, in that our results that were obtained using both first-order and second-order texture measures to better quantify heterogeneity within the pancreatic tumors. Our results suggest that homogenous features are correlated with poorer survival outcomes, in contrast to the majority of previous studies that found that increased tumoral heterogeneity on CT images is related to poorer clinical outcomes 17,[46][47][48] . Heterogeneity is a well-recognized feature of malignant tumors and presumably reflects alterations in the tissue microenvironment due to cell infiltration, angiogenesis, necrosis, and myxoid changes 13,48,49 . In prior studies, tumor heterogeneity measured on CT images correlated with histologic findings of an irregular, disorganized architectural distortion from angiogenesis and hypoxia in primary colorectal cancer and non-small-cell lung cancer 50,51 . However, contradictory findings were found in studies of primary 15 and metastatic colorectal cancer 52 , where texture variables representing less heterogeneity (e.g., lower entropy and standard deviations) were associated with poorer survival. Based on our study results, as well as those of the studies mentioned above (15,53), we conjecture that homogeneous texture features could represent more aggressive behavior in tumors, thereby representing higher cellular density or dense desmoplasia. Our study results therefore imply that texture analysis on pre-operative CT scans may be potentially used to identify patients who have a higher chance of recurrence after curative resection and therefore would benefit from extensive postoperative surveillance and adjuvant therapy. Moreover, multiple ongoing studies are focused on validating the benefit of neoadjuvant chemotherapy in patients with resectable or borderline resectable pancreatic cancer, although there are no data that clearly suggest improved survival with neoadjuvant chemotherapy 53 . In addition to endoscopic ultrasound or measuring serum   [54][55][56] , the ability to stratify prognosis in patients with initially resectable pancreas head cancer by performing texture analyses of routine preoperative CT images could be helpful for selecting candidates for neoadjuvant chemotherapy. Further research is warranted to confirm the correlation between texture features and clinical outcomes in a prospective, larger cohort and to determine whether the prognostic information from texture analyses could be clinically utilized for patients with pancreatic head cancer. Several limitations need to be addressed with respect to our study. First, as this study was retrospectively designed, the possibility of selection bias should be considered. Second, we did not take into account potential variables affecting tumor enhancement on the contrast-enhanced CT scans, including cardiac output, body mass, and blood volume. Third, although the texture parameters are relatively insensitive to the CT acquisition factors 57 , the use of three different types of scanners in our study might have resulted in the inherent variability of the texture features. Future studies using the same scanner and CT acquisition protocol to reduce other possible factors affecting texture analysis are required. Fourth, given that the external validation was not performed in our study, we cannot be certain that the result in our study could be applied to the external, prospectively recruited patients. Nonetheless, the LOOCV used for cross-validation was shown to strengthen the reliability of our study results. Thus, while the results of our study cannot be immediately applied to clinical practice, further prospective validation studies using large multicentre datasets are warranted. Lastly, contrary to several reports suggesting that 3-dimensional (3D) analysis would better account for tumor heterogeneity 58 , we performed a 2-dimensional (2D) quantitative tumor analysis by selecting the single axial image with largest tumor area. In addition, aside from the fact that 3D whole-tumor analysis is complex and time-consuming, recent studies have shown that there is no difference between 2D and 3D tumor analyses 52 .
Despite several limitations in our study, it is the first to investigate the association of first and second texture features with the prognosis in pancreas cancer head patients. In the era of Radiomics, the need for standardization is increasing to provide clinically relevant results. The number of patients included in our study was within the suggested value (10-15 patients per feature) to test prognostic power of texture features. Furthermore, we have provided details of methods used in the analysis and included clinically important variables in the analysis. Our study provides that texture-feature-based image analysis holds promise in predicting prognosis in pancreas head cancer patients, and that the prospective clinical studies may be needed to better delineate the potential of this approach.
In conclusion, lower average and standard deviation values from CT texture analyses are associated with poorer survival outcomes in pancreas head cancer patients who underwent curative resection. Texture analysis  features from routinely performed pre-operative CT images could be used as an independent imaging parameter for predicting the prognosis in these patients.

Materials and Methods
Seoul National University Bundang Hospital institutional review board approval was obtained for this study, and informed consent was waived. All methods were performed in accordance with the relevant guidelines and regulations.
Patients. From January 2006 to December 2014, 167 patients underwent resection for pancreas cancer in our institution. Among them, 122 patients who had a histopathologic diagnosis of ductal adenocarcinoma in the pancreas head were initially included in this study. Of these 122 patients, 27 patients were excluded from this study for the following reasons, as these factors could potentially influence the texture values: biliary stent placement along the common duct prior to CT examination (n = 15), different CT protocols (n = 11), and pancreatolith in the pancreas head area (n = 1). Additionally, 7 patients were excluded because their pancreas head cancers were not identifiable on the initial CT images. Finally, 88 patients were included as the sample group for our study (Fig. 3). None of these included 88 patients had undergone either preoperative radiation or chemotherapy.
CT imaging protocol. All patients underwent preoperative contrast-enhanced CT imaging with a pancreas protocol. After the acquisition of non-contrast images, iopromide, an intravenous contrast material (Ultravist 370; Bayer, Berlin, Germany) was injected via the antecubital vein using a power injector (Stellant D; Medrad, Indianola, PA) at a dose of 2 mL per kilogram of body weight at a rate of 3 mL/sec. CT scans of the pancreatic and portal venous phase were initiated after the bolus contrast media injection with delays of 20 and 60 seconds after aortic enhancement of 150 HU, respectively. Non-contrast and pancreatic phase images were acquired from the diaphragm to the umbilicus level, and portal venous phase images were obtained from the diaphragm to symphysis pubis level. Images were acquired with 16-(n = 35), 64-(n = 39), or 128-(n = 14) multi-detector CT scanners (Mx 8000, Brilliance 64, iCT256; Philips Medical Systems, Cleveland, OH). The scanning parameters were as follows: 16 × 1.5, 64 × 0.625, or 128 × 0.625 mm collimation; a rotation speed of 0.5 s; a pitch of 1.25, 0.641, or 0.993; a kvP of 120. Effective mAs ranged from 72 to 385 mAs using an automatic tube current modulation technique (Dose-Right; Philips Medical Systems). The CT images were reconstructed using filtered back projection with 4-mm thick sections at 3-mm increments.
Quantitative texture analysis. The pancreatic-phase CT images were retrieved from the picture archiving and communication system and transferred to an independent workstation for further texture analysis using software built in-house. After selecting the single axial pancreatic-phase CT image 41 showing the largest cross-sectional area of the pancreas head cancer, a polygonal region of interest (ROI) was manually drawn as large as possible within the tumor border with the consensus of two radiologists (K.Y.H. and Y.G.B., with 20 and 3 years of experience in abdominal imaging, respectively) who were blinded to the pathologic and clinical outcomes (Fig. 4). Particular attention was paid to avoiding the peripancreatic vessels while delineating the ROIs for each case. Areas of air and fatty tissues were removed from the analyses by excluding any pixels with attenuation values less than 0 Hounsfield units. Although the contouring was performed on the pancreatic-phase CT images, the portal-venous-phase CT or magnetic resonance imaging (MRI) scans were also reviewed to check whether the ROIs were accurately drawn. The median tumor areas and the number of pixels in the ROIs for the texture analyses were 132.8 cm 2 (range, 61.3 to 597.6) and 433.3 (range, 164 to 1,685), respectively.
Laplacian of the Gaussian band-pass filter was applied to detect intensity changes within the images smoothened by Gaussian distribution based on the filter sigma value 17,59 . This resulted in the images displaying features at different scales (from fine to coarse textures) associated with filter sigma values within the ROI around the pancreas head cancer. The scale was determined by filter sigma values of 1.0 (fine texture, filter width 4 pixels), 1.5 to 2.0 (medium texture, filter width 6-10 pixels), and 2.5 (coarse texture, filter width 12 pixels) 17 . The degree of image smoothening was proportional to the filter value: a higher (or lower) filter value enabled the extraction of a coarse (or fine) texture by smoothening the images to a greater (or lesser) degree 59 . The distributions of pixel values of the gray-level histograms within the ROIs were characterized by average (mean intensity of the gray-level distribution), standard deviation (the degree of dispersion), kurtosis (flatness of the histogram), and skewness (asymmetry of the histogram). Texture parameters, including angular second moment (or energy; uniformity in gray-level distribution), entropy (randomness of pixel distribution), correlation (measurement of gray-level linear dependence), and contrast (measurement of local variations) were calculated by GLCM, which represents the spatial dependence relationship between groups of neighboring pixel intensity values 13,14,45 . In general, a higher standard deviation of the pixel distribution, a higher kurtosis, a positive or negative skewness, a higher entropy, and a higher contrast and lower angular second moment and correlation represented increased heterogeneity 14-17,45 . Review of pathologic and clinical follow-up data. The pathologic and clinical follow-up data were reviewed by one radiologist (Y.J.L., with 9 years of experience in abdominal imaging). The final histopathologic reports of the surgically excised specimens were also reviewed for tumor size, presence of lymph node metastasis, resection margin involvement, and pathologic differentiation according to the 7 th American Joint Committee on Cancer staging system 60 . The pathologic results were dichotomized as follows: smaller than 2.5 cm or larger than or equal to 2.5 cm for size, positive or negative for lymph node metastasis, positive (R1) or negative (R0) for surgical margins, and well to moderately or poorly differentiated pathologic differentiation 61 . After surgery, all patients underwent clinical follow-up according to our institutional protocol, including serum cancer antigen (CA) 19-9 measurement and CT examinations at 3-to 6-month intervals. Medical records and CT examinations following surgical resection were reviewed, focusing on the presence and date of tumor recurrence or death and last follow-up date. Tumor recurrence was determined by the presence of locoregional recurrence or distant metastasis documented on a patient's medical record based on physical examination, laboratory findings, follow-up imaging studies, and pathologic reports of biopsy samples, if available. Then the patients were classified into recurrence and non-recurrence groups. DFS was defined as the period from resection to the diagnosis of the tumor recurrence or to any cause of death. The final data were collected on March 31, 2017. Patients without recurrence on the date of the most recent follow-up were censored in the analysis. Statistical analysis. The clinicopathologic results and CT texture features were compared between the recurrence and non-recurrence groups. The univariate analysis for categorical variables was performed using the chi-square test. A Mann-Whitney U test was performed to compare the continuous variables between the two groups. To dichotomize the texture features with or without filters for the survival analysis, the optimal cut-off values were determined by the value which maximizes the sum of sensitivity and specificity on a receiver-operating characteristic (ROC) curve analysis. To improve the power of prediction, additional cross-validation of the results using Leave-one-out cross validation (LOOCV) test was adopted. In LOOCV, multiple rounds of ROC analysis are carried out by using the training data and then the validation data are assigned to dichotomized group based on the cut-off point. The cut-off point selected most frequently was defined as optimized cut-off value in the LOOVC analysis. DFS was analyzed by using Kaplan-Meier method based on each of the cut-off values calculated by the ROC curve analysis, and comparisons of the dichotomized variables between groups were performed by a log-rank test. Additionally, LOOCV cross-validated Kaplan-Meier analysis was performed (Table 3). A multivariate Cox proportional hazards model with a hierarchical forward step-wise procedure was used to assess whether the texture features with or without filters were independently and significantly associated with DFS. Variables with P values less than 0.05 in the univariate Kaplan-Meier analysis were entered into a multivariate Cox proportional hazards model. All analyses were performed using SPSS version 14.0 (Chicago, IL) and Medcalc version 12.1.4.0 (Medcalc Software, Ostend, Belgium). P values less than 0.05 were considered to be statistically significant.