Pancreatic neuroendocrine tumor: prediction of the tumor grade using CT findings and computerized texture analysis

Abstract

Background

Pancreatic neuroendocrine tumors (PNET) include heterogeneous tumors with a variable degree of inherent biologic aggressiveness represented by the histopathologic grade. Although several studies investigated the computed tomography (CT) characteristics which can predict the histopathologic grade of PNET, accurate prediction of the PNET grade by CT examination alone is still limited.

Purpose

To investigate the important CT findings and CT texture variables for prediction of grade of PNET.

Material and Methods

Sixty-six patients with pathologically confirmed PNETs (grade 1 = 45, grades 2/3 = 21) underwent preoperative contrast-enhanced CT. Two reviewers determined the presence of predefined CT findings. CT texture was also analyzed on arterial and portal phase using both two-dimensional (2D) and three-dimensional (3D) analysis. Multivariate logistic regression analysis was performed in order to identify significant predictors for tumor grade.

Results

Among CT findings and CT texture variables, the significant predictors for grade 2/3 tumors were an ill-defined margin (odds ratio [OR] = 7.273), lower sphericity (OR = 0.409) on arterial 2D analysis, higher skewness (OR = 1.972) and lower sphericity (OR = 0.408) on arterial 3D analysis, lower kurtosis (OR = 0.436) and lower sphericity (OR = 0.420) on portal 2D analysis, and a larger surface area (OR = 2.007) and lower sphericity (OR = 0.503) on portal 3D analysis (P < 0.05). Diagnostic performance of texture analysis was superior to CT findings (AUC = 0.774 vs. 0.683).

Conclusion

CT is useful for predicting grade 2/3 PNET using not only the imaging findings including an ill-defined margin, but also the CT texture variables such as lower sphericity, higher skewness, and lower kurtosis.

Keywords

Abdomen/gastrointestinal computed tomography (CT)pancreas adults primary neoplasms

Introduction

Pancreatic neuroendocrine tumors (PNETs) are relatively uncommon and account for approximately 2% of all pancreatic neoplasms (1). Although PNETs are regarded as potentially malignant tumors, PNETs include heterogeneous tumors with a variable degree of inherent biologic aggressiveness represented by the histopathologic grade (1,2). According to the World Health Organization (WHO) classification, PNETs are categorized into three grades based on the Ki-67 index and the mitotic number, and which provide different prognoses (3). Contrast-enhanced computed tomography (CT) has become a mainstay for the evaluation of a variety of pancreatic diseases due to its high spatial and temporal resolution (4). Although several studies investigated the CT characteristics, which can predict the histopathologic grade of PNET, such as poor enhancement, an ill-defined margin, and pancreatic ductal dilatation, accurate prediction of the PNET grade by CT examination alone is still limited.

On the other hand, texture analysis refers to mathematical methods used to analyze the attenuation value of each voxel and their distribution within the region of interest (ROI) in order to provide a measure of intra-lesional heterogeneity (5,6). A major advantage of texture analysis is that it can quantify and maximize information obtained from CT images and which may not be seen by the naked eye of a radiologist; it thus has the potential to improve lesion characterization. According to previously published reports, CT or magnetic resonance (MR) texture analysis was helpful for differentiating benign from malignant breast lesions (7 –9), for discriminating brain metastasis from cerebral glioma, and high-grade from low-grade glioma (10 –12), for characterizing part-solid, ground-glass nodules in the lung (13,14), for differentiating metastatic lymph nodes and predicting survival and response to chemotherapy in colorectal cancer (CRC) patients, for differentiating different gastric tumors, and for discriminating different renal tumors (15,16).

However, to the best of our knowledge, there has been no published report regarding the usefulness of CT texture analysis for predicting the histopathologic grade of PNET. Therefore, the purpose of our study is to investigate the important features for predicting the histopathologic grade of PNET using CT image findings and CT texture analysis.

Material and Methods

Our institutional review board approved this study and informed patient consent was waived due to the retrospective nature of the study.

Study population

One author searched our institution’s electronic medical records and identified 101 patients who were pathologically confirmed between January 2010 and August 2014 as having PNET. Among them, 13 patients who did not undergo preoperative, contrast-enhanced CT examination including both arterial and portal phase images were excluded from our study. Subsequently, 22 additional patients who underwent CT examination on a CT scanner performing less than ten examinations were excluded in order to reduce the influence of inter-scanner difference. Finally, a total of 66 patients comprised our study population (31 men, 35 women; age range = 18–82 years; mean age = 58 years). Among these patients, 21 patients have been previously reported (17). The following surgical procedures were performed on 64 patients: pancreaticoduodenectomy (n = 3); pylorus-preserving pancreaticoduodenectomy (n = 21); distal pancreatectomy (n = 30); and enucleation (n = 10). Two patients underwent only needle biopsy. Fig. 1 illustrates a flow chart of patient enrollment.

Fig. 1.

Flow diagram of the included patients.

CT examination

Contrast-enhanced CT examinations were performed using one of the following multi-detector CT scanners: Brilliance 64-MDCT (Philips Medical Systems, Cleveland, OH, USA, n = 29); Somatom Definition dual-source CT (Siemens Medical Solutions, Forchheim, Germany, n = 13); and Sensation 16-MDCT (Siemens Medical Solutions, n = 24). The following parameters were used for the CT examinations: detector configurations = 64 × 0.6–0.625 and 16 × 0.75 mm for the 64- and 16-channel multi-detectors; table speeds = 12.0 and 46.8 mm per rotation; gantry rotation times = 0.5–0.75 s; a tube voltage = 120 kVp; tube currents = 150–200 mAs; section thicknesses = 2.5–3.2 mm; and reconstruction intervals = 2–3 mm. For contrast-enhanced studies, a total of 1.5 mL of non-ionic contrast agent (iopromide, Ultravist 370; Bayer Schering Pharma AG, Berlin, Germany) per kilogram of patient body weight was injected using a power injector (Multilevel CT, Medrad, Pittsburgh, PA, USA), at a rate of 3 mL/s via an intravenous catheter. A 20-mL flush of sterile saline followed. For determining the scanning time, the automatic bolus tracking technique was used. The ROI was placed over the abdominal aorta, and arterial phase scanning was performed after reaching a triggering threshold of 100 HU of aortic blood. The mean time delay was 22–24 s for the early arterial phase, 37–45 s for the arterial phase, and 70 s for the portal phase imaging.

CT image analysis

Two attending abdominal radiologists (JHK and MHY with 17 and seven years of clinical experience in abdominal imaging, respectively), who were blinded to the patients’ clinical information and histopathologic results, reviewed the CT images. Initially, the following CT findings were independently determined by each reviewer: (i) margin (well-defined versus ill-defined); (ii) the presence of cystic degeneration; (iii) the presence of calcification; (iv) the enhancement pattern (homogenous versus heterogeneous); (v) the enhancement degree (high-, iso-, or low-attenuation compared with the pancreatic parenchyma) seen on the pre, arterial, and portal phase; and (vi) the presence of ductal dilatation. The third reviewer (JKH with 32 years of clinical experience in abdominal imaging) determined the CT findings if there was any discrepancy between the other two radiologists.

CT texture analysis

The lesion segmentation and quantification of the texture features were performed by two independent radiologist (TWC and JHK with three and 17 years of clinical experience in abdominal imaging, respectively) using an in-house software program (Medical Imaging Solution for Segmentation and Texture Analysis) as follows: first, arterial and portal phase, axial CT images covering PNET were transferred and stored from the picture archiving and communication system (INFINITT PACS, INFINITT Healthcare, Seoul, Republic of Korea) of our hospital to a personal computer for texture analysis. Thereafter, manual segmentation of the tumor was performed on arterial phase images and portal phase images, respectively, using both the two-dimensional (2D) and three-dimensional (3D) methods. Initially, an axial image with the largest cross-sectional area of the tumor was selected and 2D segmentation was performed by manually drawing a ROI around the tumor outline. 3D segmentation was performed by determining ROIs using the same method on every axial CT image showing the lesion in order to cover the entire tumor volume (Fig. 2). After a tumor was segmented, the following texture parameters were automatically calculated: the histogram parameters including (i) mean attenuation, (ii) standard deviation, (iii) skewness, (iv) kurtosis, (v) entropy, and (vi) homogeneity; the volumetric parameters including (vii) volume, (viii) effective diameter, and (ix) surface area; the morphologic parameters including (x) sphericity and (xi) discrete compactness; and second-order texture features based on gray-level co-occurrence matrices (GLCM) which characterize the spatial distribution of gray levels in images, including (xii) GLCM moments, (xiii) GLCM angular second moment (ASM), (xiv) GLCM inverse difference moment (IDM), (xv) GLCM contrast, and (xvi) GLCM entropy.

Fig. 2.

Texture analysis software program. Manual segmentation of the tumor was performed on the working area with the long dashed red line, and segmented tumor was presented as the green-colored area. The area lined with simple red line shows a stack of consecutive axial images of CT scan. In 2D analysis (a), manual segmentation was performed on an axial image with the largest cross-sectional area of the tumor whereas tumor segmentation was performed on every axial CT images showing the lesion to cover the entire tumor volume in the 3D method (b). After a tumor was segmented, the texture parameters were automatically calculated by the program and demonstrated on the area outlined by a short dashed red line.

Statistical analysis

Each of the different CT findings between grade 1 and grade 2/3 PNETs was compared using Fisher’s exact test and the Mann–Whitney U test. For texture parameters, the Shapiro–Wilk test was performed to confirm that the variables were normally distributed. The independent sample t-test was used for normally distributed variables, otherwise the Mann–Whitney U test was used. Inter-observer agreement for CT image analysis and CT texture analysis was calculated using kappa statistics and intraclass correlation coefficient (ICC), respectively. Kappa and ICC values ≤ 0.40 indicated poor agreement; 0.41–0.60, moderate agreement; 0.61–0.80, substantial agreement; and > 0.80, almost perfect agreement. Logistic regression analysis was performed to identify the significant predictors for grade 2/3 tumors. Initially, univariate analysis was performed for each variable including the CT finding and texture parameter, and only variables with P values < 0.05 on univariate analysis were selected as input variables for multivariate analysis. The backward stepwise elimination method was used in multiple logistic regression analysis. Adjusted odds ratios (OR) per one standard deviation change were used for the texture parameters. The discriminating performance of the predictive models from multiple logistic regression analysis for grade 2/3 tumors was determined using receiver operating characteristic (ROC) analysis. To evaluate the influence of inter-scanner difference on the texture analysis, the texture parameters of PNETs were compared between CT scanners using one-way analysis of variance (ANOVA). For all statistical analyses, the SPSS 21.0 software package (SPSS Inc., Chicago, IL, USA) was used and P values < 0.05 indicated a statistically significant difference.

Results

The histopathologic results showed that of 66 PNETs, 45 tumors (68.2%) were grade 1, 16 (24.2%) were grade 2, and five (7.6%) were grade 3. Among these 66 PNETs, 47.0% (31/66) were found at the head of pancreas, 21.2% (14/66) at the body, and 31.8% (21/66) at the tail. The mean ± standard deviation of the longest diameter of all of the PNETs was 2.98 ± 2.21 cm. The mean diameter of grade 1 PNETs (2.49 ± 1.29 cm) was significantly smaller than that of grade 2/3 PNETs (4.05 ± 3.17 cm, P = 0.045).

CT findings according to the tumor grade

PNETs of grade 2/3 were more likely to show an ill-defined margin (grade 1 = 5/45, 11.1% versus grade 2/3 = 10/21, 47.6%; P = 0.003), iso-to-hypo enhancement on the arterial (grade 1 = 12/45, 26.7% versus grade 2/3 = 11/21, 52.4%; P = 0.020) and portal phases (grade 1 = 15/45, 33.3% versus grade 2/3 = 13/21, 61.9%; P = 0.016), and ductal dilation (grade 1 = 10/45, 22.2% versus grade 2/3 = 10/21, 47.6%; P = 0.047) compared with PNETs of grade 1 (Figs 3 and 4). Inter-observer agreement showed substantial to almost perfect agreement between the two reviewers (0.756–1.000). Table 1 summarizes the CT findings of PNETs according to the tumor grade.

Fig. 3.

Grade 1 PNET in an 82-year-old woman. (a) Contrast-enhanced CT image obtained during the arterial phase shows an approximately 1.3-cm sized well-circumscribed tumor (arrow) in the head of pancreas. The tumor shows strong enhancement. (b) The screenshot image shows 3D tumor segmentation on arterial phase CT images by using texture analysis software program. (c, d) Histogram of the tumor in the arterial and portal phases, respectively. Skewness of the histogram analyzed on arterial phase CT images using 3D method was −0.762 and kurtosis on 2D portal phase analysis was 0.479. The sphericity of the tumor on arterial phase 3D analysis was 0.847.

Fig. 4.

Grade 3 neuroencocrine carcinoma of the pancreas in a 39-year-old man. (a) Axial, contrast-enhanced CT image obtained during arterial phase demonstrates an approximately 4.0-cm sized ill-defined subtle low attenuated mass (arrow) in the head of pancreas. (b) On an axial, portal phase CT image covering the pancreatic body portion reveals dilatation of main pancreatic duct (arrowheads) and distal common bile duct (arrow). (c) The screenshot image shows 3D segmentation on arterial phase CT images by using texture analysis software program. (d, e) Histogram of attenuation of the tumor on arterial and portal phase, respectively. Skewness of the histogram on 3D arterial phase analysis was 0.021 and kurtosis of the histogram on 2D portal phase analysis was 0.011. In addition, the sphericity of the tumor analyzed by 3D segmentation method was 0.686.

Table 1.

CT findings of PNET according to tumor grade.

CT findings	Tumor grade				P value	kappa
CT findings	G1 (n = 45)	G2 (n = 16)	G3 (n = 5)	Total	P value	kappa
Margin
Well-circumscribed	40	10	1	51	0.003	0.909
Ill-defined	5	6	4	15
Cystic degeneration	14	3	0	17	0.227	0.756
Calcification	3	0	1	4	1.000	1.000
Enhancement
Homogeneous	22	6	2	30	0.440	0.908
Heterogeneous	23	10	3	36
Enhance patterns
Pre
Low	16	6	1	23	0.906	0.792
Iso	26	8	4	38
High	2	1	0	3
NA	1	1	0	2
Arterial
Low	6	5	4	15	0.020	0.813
Iso	6	1	1	8
High	33	10	0	43
Venous
Low	6	5	3	14	0.016	0.776
Iso	9	4	1	14
High	30	7	1	38
Duct dilation	10	7	3	20	0.047	0.818

CT texture analysis according to the tumor grade

The texture parameters of grade 1 and grade 2/3 PNETs are summarized in Suppl. Table 1. In arterial 2D and arterial 3D analysis, grade 2/3 tumors showed a significantly lower mean attenuation, higher skewness, a larger volume, a larger effective diameter, lower sphericity, lower GLCM moments, and higher GLCM IDM compared to grade 1 tumors (P < 0.05). In addition, grade 2/3 tumors showed a significantly higher GLCM entropy in arterial 2D analysis and a larger surface area in arterial 3D analysis than grade 1 tumors (P < 0.05, Figs 3 and 4). Portal 2D and portal 3D analysis revealed that grade 2/3 tumors are more likely to show higher skewness, lower kurtosis, higher homogeneity, larger volume, larger effective diameter, lower sphericity, and lower GLCM moments than grade 1 tumors (P < 0.05). In addition, in portal 3D analysis, grade 2/3 tumors showed a significantly larger surface area compared with grade 1 tumors (P < 0.05). Inter-observer agreement for texture parameters showed substantial to almost perfect agreement (0.646–0.997) except for surface area and discrete compactness in the arterial and portal 2D analysis.

Important features for the prediction of tumor grade

The significant predictors for grade 2/3 tumor were as follows: an ill-defined lesion margin (OR = 7.273, P = 0.002); lower sphericity (adjusted OR = 0.409, P = 0.009, ICC = 0.704) on the 2D arterial phase; higher skewness (adjusted OR = 1.972, P = 0.035, ICC = 0.915) and lower sphericity (adjusted OR = 0.408, P = 0.005, ICC = 0.845) on the 3D arterial phase; lower kurtosis (adjusted OR = 0.436, P = 0.033, ICC = 0.679) and lower sphericity (adjusted OR = 0.420, P = 0.015, ICC = 0.687) on the 2D portal phase; and a larger surface area (adjusted OR = 2.007, P = 0.035, ICC = 0.997) and lower sphericity (adjusted OR = 0.503, P = 0.037, ICC = 0.883) on the 3D portal phase (Table 2). The area under the curve (AUC) for the prediction of grade 2/3 tumor were 0.683 for the CT finding analysis, 0.743 for the arterial 2D analysis, 0.774 for the arterial 3D analysis, 0.767 for the portal 2D analysis, and 0.745 for the portal 3D analysis. The diagnostic performance of the CT texture analysis using the 3D arterial phase (AUC = 0.774) for predicting grade 2/3 tumor was superior to that of the CT findings (AUC = 0.683), although without a significant difference (P = 0.151). In the comparison of the texture analyses between the CT scanners (Appendix E1, E2, online only), there was no texture feature which showed a significant difference between the CT units (P > 0.05), except for GLCM IDM in the arterial 2D (P = 0.045) and arterial 3D analysis (P = 0.048) and sphericity in the portal 3D analysis (P = 0.019).

Table 2.

Multivariate analyses of important features for prediction of tumor grade.

	Characteristics	OR	95% CI	P value
CT finding	Ill-defined margin	7.273	2.055–25.742	0.002
Texture Analyses
Arterial phase
2D	Sphericity*	0.409	0.208–0.803	0.009
3D	Skewness*	1.972	1.050–3.705	0.035
3D	Sphericity*	0.408	0.219–0.762	0.005
Portal phase
2D	Kurtosis*	0.436	0.203–0.936	0.033
2D	Sphericity*	0.420	0.208–0.848	0.015
3D	Surface area (mm²)*	2.007	1.050–3.835	0.035
3D	Sphericity*	0.503	0.264–0.959	0.037

Data are adjusted odds ratios (ORs) per one standard deviation change.

Discussion

Our study demonstrated that an ill-defined margin of the lesion (OR = 7.273, P = 0.002), lower sphericity (adjusted OR = 0.408–0.503, P < 0.05), higher skewness (adjusted OR = 1.972, P = 0.035) seen on the 3D arterial phase; lower kurtosis (adjusted OR = 0.436, P = 0.033) on the 2D portal phase; and a larger surface area (adjusted OR = 2.007, P = 0.035) seen on the 3D portal phase were significant predictors for grade 2/3 tumors. The diagnostic performance of computerized, whole-tumor, volumetric, 3D texture analysis using arterial phase images (AUC = 0.774) was better than the CT findings (AUC = 0.683) for predicting grade 2/3 PNETs, although there was not a significant statistical difference (P = 0.151).

There have been several studies which investigated the CT characteristics which can predict the grade of PNET. Kim et al. reported that a portal enhancement ratio <1.1, poorly defined margin, tumor size >3 cm, bile duct dilatation, and vascular invasion were significant predictors for grade 3 PNET distinguished from grade 1/2 tumors (18). d’Assihnies et al. found that tumor blood flow seen on perfusion CT is correlated with the histoprognostic findings including the proliferative index and tumor grade according to the WHO classification (19). Although data acquisition was performed at a single axial plane without volumetric perfusion measurement, this study also provided radiopathologic correlation by assessing the tumor microvascular density using histologic techniques. In our study, grade 2/3 tumors more frequently showed an ill-defined margin (OR = 7.273, P = 0.002) than grade 1 tumors, and which parallels the results of previous reports (17,18,20). In addition, grade 1 tumors showed stronger enhancement than higher grade tumors. This result is also well coordinated with the previous report by Kim et al. which demonstrated that grade 1 PNETs are enhanced more prominently than higher grade tumors on MR imaging (MRI) (20). d’Assihnies et al. also showed that the lower grade PNETs show significantly increased tumor blood flow than higher grade lesions by analyzing the perfusion CT data (19). In our study, the mean diameter of grade 2/3 PNETs (4.05 ± 3.17 cm) was significantly larger than that of grade 1 PNETs (2.49 ± 1.29 cm, P = 0.045) and the sphericity of grade 2/3 tumors (adjusted OR = 0.408–0.503, P < 0.05) was significantly lower than that of grade 1 tumors.

CT texture analysis is a useful tool for the quantitative assessment of tumor heterogeneity (5). It is important to assess intratumoral heterogeneity as tumors with high intratumoral heterogeneity have shown a poorer prognosis or higher histopathologic grade and which reflects the intrinsic biologic aggressiveness, as noted in several published reports (6,10 –12). However, the increased heterogeneity of a tumor is not always associated with increased biologic aggressiveness or a poorer prognosis. Ng et al. reported that homogenous CT texture features of primary CRC are associated with a poorer prognosis in CRC patients, and the authors suggested that these homogenous texture features may be related to the vascular permeability of the tumor (21). In our study, grade 2/3 PNETs showed higher GLCM IDM, higher GLCM entropy, lower kurtosis, and higher skewness compared with grade 1 tumors, and which, in general, represents decreased heterogeneity (21,22). Therefore, more validation studies will be required for clinical practice. In addition, the presence of cystic degeneration and calcification of PNETs may also have contributed to the histogram parameters such as kurtosis and skewness. In our study, grade 1 PNETs more often showed cystic degeneration (31.1%, 14/45) than grade 2/3 tumors (14.3%, 3/21), although statistical significance was not found (P = 0.227). This higher proportion of cystic change in grade 1 tumors than in grade 2/3 lesions is not contrary to earlier observations in previously published reports (20,23). On the other hand, calcifications seen on preoperative CT images were reported in 10.8–22.7% of PNETs in previously published literature and were seen to be more common in grade 2 PNETs than in grade 1 PNETs (18,24,25). However, in our study, only 6.7% (3/45) of grade 1 tumors and 4.8% (1/21) of grade 2/3 tumors showed calcification on CT images (P = 1.000). Therefore, we speculate that in our study population, the presence of cystic degeneration may have had a greater impact on the histogram parameters of texture analysis than the presence of calcification. The lower skewness with a more negative value of grade 1 PNETs compared with grade 2/3 tumors may be explained by the higher mean attenuation of the enhancing tumor portion as well as the more frequently found low attenuation voxels caused by cystic degeneration which appears as a longer or fatter tail on the left side of the histogram. On the other hand, the distribution with positive kurtosis indicates that it has heavier tails and a higher peak than the normal distribution, whereas a distribution with negative kurtosis has lighter tails and a flatter peak (26). Therefore, the higher kurtosis of grade 1 tumors than that of grade 2/3 tumors, as seen on portal phase images, may also be explained by the heavier tail of grade 1 tumors due to the more frequent cystic degeneration.

There are several limitations to our study. First of all, as it was of retrospective design, there may have been a selection bias. Second, our study included only a small number of patients, and, in particular, there were only five patients with grade 3 PNETs. A comparison between grade 1/2 tumors and grade 3 tumors could not be performed due to the insufficient number of grade 3 tumors. Third, the CT scanners used in our study were not uniform among all of the patients, and, consequently, an inter-scanner difference may have increased the variability of the lesion attenuation as well as affecting the texture analysis results. Therefore, in order to evaluate the influence of inter-scanner differences on texture analysis, the texture parameters of PNETs were compared between the CT scanners, and the results showed that except for GLCM IDM on the arterial phase analysis and sphericity on the 3D portal phase analysis, there was no texture parameter which showed a significant difference between the CT scanners. However, further studies with a larger study population and uniform CT scanner and examination protocol are warranted in order to validate the findings of our study. Finally, in our study, tumor segmentation was manually performed by the radiologist, and it can be influenced by a subjective tendency or bias. To assess the inter-observer variability for texture parameters, a second radiologist re-segmented all images independently and the results showed substantial to almost perfect agreement for the majority of texture parameters. Nevertheless, we believe that robust automatic boundary extraction method should be further developed in order to address the variability issue.

In conclusion, CT is helpful for the prediction of grade 2/3 PNETs using not only imaging findings, including an ill-defined margin, iso-to-hypo enhancement, and ductal dilation, but also CT texture parameters such as lower sphericity, higher skewness, and lower kurtosis.

Footnotes

Acknowledgments

The authors thank Bonnie Hami, MA, USA for her editorial assistance in the preparation of this manuscript.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

References

Reid

Balci

Saka

et al.

Neuroendocrine tumors of the pancreas: current concepts and controversies. Endocr Pathol 2014; 25: 65–79.

Klimstra

Modlin

Coppola

et al.

The pathologic classification of neuroendocrine tumors: a review of nomenclature, grading, and staging systems. Pancreas 2010; 39: 707–712.

Bosman

Carneiro

Hruban

et al.

WHO classification of tumours of the digestive system, 4th ed. Lyon: International Agency for Research on Cancer, 2010.

Bashir

Gupta

. MDCT evaluation of the pancreas: nuts and bolts. Radiol Clin North Am 2012; 50: 365–377.

Ganeshan

Miles

. Quantifying tumour heterogeneity with CT. Cancer Imaging 2013; 13: 140–149.

Davnall

Yip

Ljungqvist

et al.

Assessment of tumor heterogeneity: an emerging imaging tool for clinical practice?

Insights Imaging 2012; 3: 573–589.

Gibbs

Turnbull

. Textural analysis of contrast-enhanced MR images of the breast. Magn Reson Med 2003; 50: 92–98.

Chen

Giger

et al.

Volumetric texture analysis of breast lesions on contrast-enhanced magnetic resonance images. Magn Reson Med 2007; 58: 562–571.

Holli

Laaperi

Harrison

et al.

Characterization of breast cancer types by texture analysis of magnetic resonance images. Acad Radiol 2010; 17: 135–141.

10.

Ryu

Choi

Park

et al.

Glioma: application of whole-tumor texture analysis of diffusion-weighted imaging for the evaluation of tumor heterogeneity. PLoS One 2014; 9: e108335–e108335.

11.

Skogen

Ganeshan

Good

et al.

Measurements of heterogeneity in gliomas on computed tomography relationship to tumour grade. J Neurooncol 2013; 111: 213–219.

12.

Zacharaki

Wang

Chawla

et al.

Classification of brain tumor type and grade using MRI texture and shape in a machine learning scheme. Magn Reson Med 2009; 62: 1609–1618.

13.

Lee

Goo

et al.

Usefulness of texture analysis in differentiating transient from persistent part-solid nodules(PSNs): a retrospective study. PLoS One 2014; 9: e85167–e85167.

14.

Chae

Park

et al.

Computerized texture analysis of persistent part-solid ground-glass nodules: differentiation of preinvasive lesions from invasive pulmonary adenocarcinomas. Radiology 2014; 273: 285–293.

15.

Ba-Ssalamah

Muin

Schernthaner

et al.

Texture-based classification of different gastric tumors at contrast-enhanced CT. Eur J Radiol 2013; 82: e537–543.

16.

Yan

Liu

Wang

et al.

Angiomyolipoma with minimal fat: differentiation from clear cell renal cell carcinoma and papillary renal cell carcinoma by texture analysis on CT images. Acad Radiol 2015; 22: 1115–1121.

17.

Kim

Eun

Kim

et al.

Pancreatic neuroendocrine tumour (PNET): Staging accuracy of MDCT and its diagnostic performance for the differentiation of PNET with uncommon CT findings from pancreatic adenocarcinoma. Eur Radiol 2016; 26: 1338–1347.

18.

Kim

et al.

Neuroendocrine neoplasms of the pancreas at dynamic enhanced CT: comparison between grade 3 neuroendocrine carcinoma and grade 1/2 neuroendocrine tumour. Eur Radiol 2015; 25: 1375–1383.

19.

d’Assignies

Couvelard

Bahrami

et al.

Pancreatic endocrine tumors: tumor blood flow assessed with perfusion CT reflects angiogenesis and correlates with prognostic factors. Radiology 2009; 250: 407–416.

20.

Kim

Eun

Kim

et al.

Staging accuracy of MR for pancreatic neuroendocrine tumor and imaging findings according to the tumor grade. Abdom Imaging 2013; 38: 1106–1114.

21.

Ganeshan

Kozarski

et al.

Assessment of primary colorectal cancer heterogeneity by using whole-tumor texture analysis: contrast-enhanced CT texture as a biomarker of 5-year survival. Radiology 2013; 266: 177–184.

22.

Albregtsen F. Statistical texture measures computed from gray level cooccurrence matrices. Technical Note. Oslo: Department of Informatics, University of Oslo, 1995.

23.

Gallotti

Johnston

Bonaffini

et al.

Incidental neuroendocrine tumors of the pancreas: MDCT findings and features of malignancy. Am J Roentgenol 2013; 200: 355–362.

24.

Rodallec

Vilgrain

Couvelard

et al.

Endocrine pancreatic tumours and helical CT: contrast enhancement is correlated with microvascular density, histoprognostic factors and survival. Pancreatology 2006; 6: 77–85.

25.

Poultsides

Huang

Chen

et al.

Pancreatic neuroendocrine tumors: radiographic calcifications correlate with grade and metastasis. Ann Surg Oncol 2012; 19: 2295–2303.

26.

DeCarlo

. On the Meaning and Use of Kurtosis. Psychological Methods 1997; 2: 292–307.