Differential Diagnosis of Nonhypervascular Pancreatic Neuroendocrine Neoplasms From Pancreatic Ductal Adenocarcinomas, Based on Computed Tomography Radiological Features and Texture Analysis

ARTICLE IN PRESS

Original Investigation

Differential Diagnosis of Nonhypervascular Pancreatic Neuroendocrine Neoplasms From Pancreatic Ductal Adenocarcinomas, Based on Computed Tomography Radiological Features and Texture Analysis Haopeng Yu, MD#, Zixing Huang, MD#, Mou Li, MD, Yi Wei, MD, Lin Zhang, bachelor of medicine, Chengmin Yang, bachelor of medicine, Yongchang Zhang, bachelor of medicine, Bin Song, bachelor of medicine Abbreviations AP arterial phase CT computed tomography LASSO least absolute shrinkage and selection operator MDAC maximum diameter on axial section MRI magnetic resonance imaging PDAC pancreatic ductal adenocarcinoma PEP progressive enhancement in portal vein phase

Rationale and Objectives: To determine computed tomography (CT) radiological features and texture features that are rewarding in differentiating nonhypervascular pancreatic neuroendocrine neoplasms (PNENs) from pancreatic ductal adenocarcinomas (PDACs). Materials and Methods: We compared patients to pathologically proven nonhypervascular PNENs and age-matched controls with pathologically proven PDACs in a 1:2 ratio. Preoperative CT images in the arterial phase (AP) and portal vein phase (PVP) were obtained. Two radiologists independently reviewed the morphological characteristics of each tumor. Three-dimensional regions of interest (ROIs), drawn using ITK-SNAP software, were input into AK software (Artificial Intelligent Kit, GE) to extract texture features from AP and PVP images. Differences between PNENs and PDACs were analyzed with the chi-squared test, least absolute shrinkage and selection operator, kappa statistics, and uni- and multivariate logistic regression analyses. Results: In total, 40 nonhypervascular PNENs and 80 PDACs were evaluated. Maximum diameter on axial section, margin, calcification, vascularity in the tumor, and tumor heterogeneity were significantly different between PDACs and nonhypervascular PNENs. Multivariate analysis showed well-defined tumor margin (odds ratio: 21.0) and presence of calcification (odds ratio: 4.4) were significant predictors of nonhypervascular PNENs. The area under the receiver operating characteristic curve of the radiological feature model, AP texture model, and PVP texture model were 0.780, 0.855, and 0.929, respectively, based on logistic regression. Conclusion: A well-defined margin and calcification in the tumor were helpful in discriminating nonhypervascular PNENs from PDACs. Texture analysis of contrast-enhanced CT images could be beneficial in differentially diagnosing nonhypervascular PNENs and PDACs.

PNEN Key Words: Computed tomography; pancreatic ductal adenocarcinoma; nonhypervascular pancreatic neuroendocrine neoplasm; texture analysis.

Acad Radiol 2019; &:1–10 From the Department of Radiology, West China Hospital of Sichuan University, Guoxue Lane 37#, Chengdu 610041, PR China. Received December 21, 2018; revised May 30, 2019; accepted June 10, 2019. Address correspondence to: S.B. e-mails: [email protected], [email protected] # These authors contributed equally to this work. © 2019 The Association of University Radiologists. Published by Elsevier Inc. All rights reserved. https://doi.org/10.1016/j.acra.2019.06.012

1

ARTICLE IN PRESS Academic Radiology, Vol &, No &&, && 2019

YU ET AL

pancreatic neuroendocrine neoplasm

© 2019 The Association of University Radiologists. Published by Elsevier Inc. All rights reserved.

PVP portal vein phase TA texture analysis UPDD upstream pancreatic duct dilation

INTRODUCTION

P

ancreatic neuroendocrine neoplasms (PNENs) are rare tumors and constitute approximately 1
3% of all pancreatic neoplasms (1
3). The prevalence of PNENs is approximately 35 per 100,000 individuals in the United States of America (4), and the incidence of the disease is markedly lower in the Asia-Pacific region than in the rest of the world, possibly because of diet and/or genetic differences (1). Pancreatic neuroendocrine neoplasms can exhibit multiple clinical symptoms because of different hormones secreted by functioning PNENs; however, nonfunctioning neuroendocrine neoplasms are more frequent than their functioning counterparts (4
6). Surgery is the only curative treatment for PNENs. If complete resection is unfeasible, then a nonsurgical treatment is chosen, depending on the patient’s particular circumstances (7). Pancreatic ductal adenocarcinoma (PDAC) is one of the most aggressive malignancies. It is expected to cause more than 23,000 cancer-related deaths in America in 2019 (8). As with PNENs, surgery offers the only opportunity to cure PDACs. Adjuvant and neoadjuvant treatment are also recommended. Compared to the 5-year survival rate of PNENs, which is about 92
100% at early stage and 57% at advanced stage (9), the 5-year survival rate of PDACs is approximately 5
7%; after surgical resection, the 5-year survival rate is no more than 25% (10,11). A PNEN typically presents as a well-circumscribed solid mass that appears hyperattenuated in arterial phase (AP) images and portal venous phase (PVP) images because of a rich capillary network (2,12,13). However, 13
49% of PNEN cases may not show hyperenhancement on AP images (14
16). Such nonhypervascular PNENs may be difficult to differentiate from PDACs, which typically appear as hypoenhanced, relative to the nontumoral pancreatic parenchyma (17). In addition, therapy options for PNENs and PDACs are markedly different (7) because PNENs overall have a more indolent biology and present a more favorable long-term outcome after surgery, even with metastasis, than do PDACs (3,18,19). Thus, an accurate preoperative differentiation of these two types of tumors is of pivotal importance. Biopsy remains the gold standard of diagnosis. Endoscopic ultrasonography with fine-needle aspiration is used to obtain tissue sampling before surgery or before further treatment; however, such an invasive procedure may cause severe complications and is unnecessary before surgery (20,21). At present,

2

multiphasic CT or magnetic resonance imaging (MRI) is required before surgery, and some qualitative and quantitative features of MRI and CT may be useful in the differential diagnosis (3,14,22); however, the positive predictive values of these radiological features remain controversial. Texture analysis (TA) is a rapidly advancing method in the field of medical image analysis. It provides an objective quantitative assessment of tumor heterogeneity by analyzing the distribution of voxel gray levels, coarseness, and regularity in an image without requiring additional invasive procedures (23). Texture analysis was successfully used recently for assessing colorectal cancer, head and neck cancer, breast cancer, and lung cancer (24
27). Numerous studies have previously applied TA to pancreatic neoplasms (28
31) and have proven that TA allows differential diagnosis between malignant and benign lesions in the pancreas. Therefore, the purpose of our study was to determine the TA features and radiological features that are rewarding in the differential diagnosis of nonhypervascular PNENs and PDACs. MATERIALS AND METHODS Patients

This retrospective study was approved by the institutional review board of ZZZ Hospital, and the requirement for obtaining informed consent was waived. Between November 2008 and May 2017, 447 patients with pathologically proven PNENs were identified through a review of our pathologic databases. During this time period, we used the dual-phase CT protocol for the preoperative evaluation of pancreatic neoplasms. The inclusion criteria of the patients were as follows: (1) patients who underwent dual-phase protocol CT examinations within at least 60 days before surgery; (2) patients who had a diagnosis of nonhypervascular PNEN tumor (i.e., the degree of enhancement of the tumors was less than or close to that of the surrounding normal pancreas in the AP); and (3) patients who had not undergone local or systemic treatment before surgery. The exclusion criteria were as follows: (1) the presence of severe motion or metallic artifacts and (2) a tumor size <10 mm because background organs may affect the lesion texture measurement in TA. In addition, the same databases were used to randomly identify agematched controls with PDACs as the comparison group in a 1:2 ratio (31).

ARTICLE IN PRESS Academic Radiology, Vol &, No &&, && 2019 DIFFERENTIAL DIAGNOSIS OF NONHYPERVASCULAR PANCREATIC

Computed Tomography Examination and Imaging Analysis

On account of the retrospective design of this study, various CT scanners were used and included a 64-channel multidetector CT (MDCT) scanner (Brilliance64; Philips Medical Systems, Eindhoven, the Netherlands), 128-MDCT scanner (Somatom Definition AS+; Siemens Healthcare Sector, Forchheim, Germany), and a dual-source CT system (Somatom Definition Flash; Siemens Healthcare Sector). The scanning parameters were as follows: voltage, 120 kV; amperage, 200
210 mA; rotation time, 0.5
0.75 seconds; pitch, 0.8
1.0; and section thickness, 1.0
7.0 mm. An anionic contrast medium (Omnipaque 350; GE Healthcare, Chicago, IL), administered at a dose of 1.5 mL/kg body weight, was injected intravenously using a power injector at a rate of 3 mL/s. With the trigger threshold of the aorta reaching 100 HU, a three-phase scan was obtained. It included the precontrast phase image, the AP image (at the trigger), and the PVP image (30 seconds after the trigger). Patients were scanned with three types of CT scanners: Brilliance64 scanner (40% of including patients), Somatom Definition AS+ (15% of including patients), and Somatom Definition Flash (45% of including patients). Two abdominal radiologists with 5 years and 8 years of experience in abdominal imaging, respectively, who were unaware of the pathological diagnosis, reviewed the CT images for a consensus at a PACS workstation (Syngo-Imaging, version VB36A; Siemens Healthcare). The CT radiological features that were assessed included (1) maximum diameter on the axial section (MDAC), (2) tumor margin (well-defined vs. ill-defined), (3) presence of calcification in the tumor, (4) presence of cystic degeneration (solid vs. cystic), (5) progressive enhancement in portal vein phase (PEP), (6) presence of vascularity in the tumor, (7) tumor homogeneity in PVP (homogeneous vs. heterogeneous), (8) presence of peripancreatic vascular invasion, and (9) upstream pancreatic duct dilation (UPDD). Well-defined tumor margin: Sharp and easily distinguishable boundary between the tumor and surrounding normal tissues in precontrast phase CT images. PEP: The enhancement of tumor in the PVP was higher than that in the AP compared to surrounding normal tissues. Presence of vascularity in the tumor: In PVP or AP, blood vessels in the tumor clearly manifested itself, which maybe tumor’s supply vessels. Tumor homogeneity in PVP: Whether tumor’s parenchymal enhancement is homogeneous in PVP. Presence of peripancreatic vascular invasion: Whether tumor invades adjacent vessels and cause lumen obstruction or uneven vessel wall. UPDD: Upstream pancreatic duct diameter
5 mm.

Texture Feature Extraction

A radiologist manually delineated the regions of interest (ROIs) within a tumor (approximately 1 mm from the tumor margin) to define the three-dimensional (3D) volume area in

the AP and PVP images by using ITK-Snap software (this open source software is available at www.itk-snap.org). Blood vessels, calcification, and cystic areas in the tumor were excluded. Texture features of all patients were extracted and analyzed using A.K. software (Analysis-Kit; GE Healthcare). Three hundred eighty-five texture features of six types were extracted: Gray level histogram, run-length matrix, gray level co-occurrence matrix (GLCM), Haralick features, gray level size zone matrix, and form factor feature. The histogram indicated that the intensity of pixels and the run-length matrix depicted the amount of homogeneity in specific directions. The co-occurrence matrix and Haralick features provided information about the gray level value distribution of pixel pairs in all directions. The gray level size zone matrix is efficient for characterizing texture homogeneity, nonperiodicity, or a speckle-like texture (32,33). By using the histogram feature of the software, we extracted the texture feature parameters and generated a quantitative or qualitative description of the texture, based on the gray value of the images. The transform matrix texture reflected high-level information of the ROI by a series of matrix transformations such as gray level co-occurrence metrics and run-length metrics. With the wavelet transform, we analyzed the characteristics of the ROI through different levels of resolution. With the filter-transform texture, we obtained a series of target features by using different types of filters such as log transformation and Gaussian transformation. Statistical Analysis

Interobserver agreement of the radiological features was calculated by using weighted k statistics for categorical variables and using intraclass correlation coefficient (ICC) for continuous variables to assess feature reproducibility in repeated delineations (k value or ICC < 0 no agreement, 0
0.20 slight agreement, 0.21
0.40 fair agreement, 0.41
0.60 moderate agreement, 0.61
0.80 substantial agreement, and 0.81
1 almost perfect agreement). Owing to the large number of texture features and relatively small cohort size, redundant texture features need to be removed. Hence, all texture features were analyzed the correlation of any two texture features and redundant texture features were considered with a linear correlation coefficient greater than 0.9. The least absolute shrinkage and selection operator (LASSO) regression method (34) was then applied to avoid overfitting in the high-dimension data analysis. By optimizing the tuning parameter (λ) in the LASSO regression, most coefficients of the features were reduced to zero and the remaining features with nonzero coefficients were selected. Therefore, the most distinguishable texture features were identified (35). Each radiological feature was compared by using the chi-square test or Fisher’s exact test for the categorical variables and using Student’s t test or one-way analysis of variance for the continuous variables. Statistical significance was assumed at a confidence level of 0.05. The selected texture features and radiological features were analyzed by univariate logistic 3

ARTICLE IN PRESS Academic Radiology, Vol &, No &&, && 2019

YU ET AL

regression analysis. The features revealed as statistically significant with univariate logistic regression analysis were then analyzed with multivariate logistic regression analysis. Receiver operating characteristic curve analysis was conducted to estimate the diagnostic performance for differentiating nonhypervascular PNENs from PDACs for the regression model. All data analyses were conducted using R software version 3.3.2 (R Foundation for Statistical Computing, Vienna, Austria), SPSS version 22 (SPSS Inc., Chicago, IL), and Medcalc v.15.8 (Medcalc Software, Ostend, Belgium). RESULTS The final study population for the nonhypervascular PNENs group consisted of 40 patients (23 men and 17 women; mean age, 49.7 § 12.6 years; age range, 21
69 years; 33 nonfunctional and 7 functional neoplasms) and their tumor diameters ranged 11
179 mm (52.8 § 33.4 mm). The PDACs group consisted of 80 patients (52 men and 28 women; mean age, 50.4 § 14.1 years; age range, 30
69 years) and the tumor diameters ranged 15
85 mm (33.2 § 12.9 mm). No statistical significance was obtained from the patients’ sex (p = 0.42) and location distribution (p = 0.67), whereas the diameter of the tumors showed a statistical significant difference (p = 0.00) in the comparison of nonhypervascular PNENs and PDACs group. Age differences between the two groups were maintained to a minimum because of the age-matching enrollment protocol. The characteristics of the patients and radiological features are summarized in Table 1. The flow diagram for patient enrollment is illustrated in Figure 1. Radiological Features Evaluation

For the radiological features, the tumor margin (p = 0.00), calcification (p = 0.02), vascularity of the tumor (p = 0.01), and tumor heterogeneity (p = 0.04) were statistically significant in the nonhypervascular PNENs group and the PDACs

group. However, UPDD (p = 0.13), peripancreatic vascular invasion (p = 0.36), cystic degeneration (p = 0.05), and PEP (p = 0.22) showed no significant difference between the groups. On univariate analysis, a well-defined margin (p = 0.00, odds ratio [OR] = 21.00) and homogenous (p = 0.04, OR = 2.30) tumor with calcification (p = 0.01, OR = 4.36), and vascular involvement (p = 0.01, OR = 2.61), and that is dissociated from the peritumoral parenchyma (p = 0.00, OR = 21.00) diameter >39 mm (p = 0.00, OR = 1.04) is apt to be a PNEN rather than a PDAC. The sensitivity (Sen), specificity (Spe), OR, and 95% confidence intervals (95% confidence of interval [CI]) for each radiological feature that were significant in univariate analysis are presented in Table 2. With multivariate logistic regression analysis, only a well-defined tumor margin (OR = 14.63; 95% CI: 2.82, 75.99) and calcification in the tumor (OR = 4.00; 95% CI: 1.03, 15.59) were independent positive predictors of nonhypervascular PNENs. The two observers in our study showed almost perfect interobserver agreement in classifying the presence of calcification in the tumor (k = 0.92), MDAC (ICC = 0.94), tumor margin (k = 0.87), and the presence of vascularity in the tumor (k = 0.88). They had substantial agreement for tumor homogeneity in the PVP (k = 0.80), presence of peripancreatic vascular invasion (k = 0.74), presence of cystic degeneration (k = 0.68), and PEP (ICC = 0.69). Texture Feature Evaluation

Ten texture features of AP and 11 texture features of PVP were selected by LASSO (Fig 2). Texture features such as Quantile0.025, maximum diameter in three dimensions, cluster shade, and entropy were useful in the AP and PVP group images. Three texture features of the AP and seven texture features of the PVP remained after univariate logistic regression analysis. Two texture features of the AP and three texture features of the PVP were identified as independent

TABLE 1. Demographic Characteristics and Computed Tomography Features of Nonhypervascular PNEN and PDAC Characteristic and Features

Nonhypervascular PNEN (n = 40)

PDAC (n = 80)

p Value

Age (y)* Sex (male/female) Location (head/body-tail) Functional PNENs (yes/no) MDAC (mm)* Margin (well-defined/ill-defined) Calcification (yes/no) Cystic degeneration (solid/cystic) PEP (yes/no) Vascular present (yes/no) Homogeneous (yes/no) Peripancreatic vascular invasion (yes/no)

49.70 § 12.64 23/17 26/14 7/33 52.83 § 33.42 14/26 9/31 30/10 17/23 23/17 27/13 18/22

50.40 § 14.07 52/28 57/23
33.23 § 12.93 2/78 5/75 71/9 25/55 27/53 38/42 29/51


0.42 0.67 0.00 0.00 0.02 0.05 0.22 0.01 0.04 0.36

The data are presented as the mean § the standard deviation. MDAC, maximum diameter in axial section; PDAC, pancreatic ductal adenocarcinoma; PEP, progressively enhancement in portal vein phase; PNEN, pancreatic neuroendocrine neoplasm; PVP, portal vein phase; UPDD, upstream pancreatic duct dilation. Bold values indicated significant difference between two groups.

4

ARTICLE IN PRESS Academic Radiology, Vol &, No &&, && 2019 DIFFERENTIAL DIAGNOSIS OF NONHYPERVASCULAR PANCREATIC

Figure 1. Flow diagram of the enrolled patients *The pancreatic ductal adenocarcinoma (PDAC) group was age-matched to the nonhypervascular pancreatic neuroendocrine neoplasm (PNEN) group in a 1:2 ratio by using a random method. 4 Nonhypervascular PNENs were selected from all PNENs by the consensus of two abdominal radiologists with 15 and 13 years of experience in abdominal imaging.

variants in the differential diagnosis of nonhypervascular PNENs from PDAC by using multivariate logistic regression analysis (Table 3). In addition, multivariate logistic regression analysis revealed that the Maximum3Ddiameter of the PVP was the texture feature with the best predictive value (i.e., the area under the curve, 0.788; 95% confidence interval [CI], 0.704-0.857). The receiver operating characteristic curves of four logistic regression models were generated (Fig 3); three of these models were built independently from radiological features, AP texture features, and PVP texture features. An integrated feature model was built using all independent predictors of radiological features and texture features. The PVP texture models demonstrated the highest area under the curve at 0.926 (95% CI: 0.864
0.966) in

the differential diagnosis of nonhypervascular PNENs from PDACs. The diagnostic performance of these models is shown in Table 4. Pairwise comparison of the three models indicated that the PVP texture model was statistically significantly different from the radiological feature model (p = 0.00), AP texture model (p = 0.02), and integrated model (p = 0.03).

DISCUSSION Differentiating nonhypervascular PNENs and PDACs with CT images can be a difficult task; however, some radiological features, at radiologists’ fingertips, demonstrated usefulness for his differential diagnosis in our study. Furthermore,

TABLE 2. Sensitivity and Specificity of Significant Computed Tomography Radiological Features in Differentiating Nonhypervascular PNEN from PDAC, Based on Univariate Analysis CT Features

Sensitivity

Specificity

Odds Ratio

MDAC (>39 mm)* Well-defined margin With calcification Vascularity in the tumor Homogeneity

55.00 (38.5, 70.7) 35.00 (20.6, 51.7) 22.50 (10.8, 38.5) 57.50 (40.9, 73.0) 52.50 (41.0, 63.8)

80.00 (69.6, 88.1) 97.50 (91.3, 99.7) 93.75 (86.0, 97.9) 66.25 (54.8, 76.4) 67.50 (50.9, 81.4)

1.04 (1.02, 1.07) 21.00 (4.47, 98.61) 4.36 (1.35, 14.04) 2.61 (1.19, 5.69) 2.30 (1.04, 5.08)

Unless otherwise indicated, the data in parentheses are the 95% confidence interval. * The number in the parentheses is the threshold.MDAC, maximum diameter on axial section; PDAC, pancreatic ductal adenocarcinoma; PNEN, pancreatic neuroendocrine neoplasm.

5

ARTICLE IN PRESS YU ET AL

diagnostic performance improved with quantitative evaluation through TA. Our study showed that some morphological features of tumors on CT (i.e., a well-defined tumor, the homogeneity of the tumor, a tumor with calcification, a tumor with visible blood vessel) could be helpful for distinguishing nonhypervascular PNENs from PDACs. These features can be explained by the pathological nature of these tumors. Pancre-

Academic Radiology, Vol &, No &&, && 2019

atic neuroendocrine neoplasms show higher cellularity and less surrounding infiltration than do PDACs (36,37). Most cases of nonhypervascular PNENs in our study were nonfunctional. These tumors are usually larger in the preliminary diagnosis because of the lack of clinical symptoms and because PNENs possess relatively indolent biological behavior, compared to PDACs. This finding explains why nonhypervascular PNENs were larger than PDACs, based on the

Figure 2. The LASSO process for selecting features extracted from the AP (a) and the PVP (b). AP, arterial phase; LASSO, least absolute shrinkage and selection operator; PVP, portal vein phase Note: The tuning parameter (λ) is calculated in the LASSO regression via the minimum criteria. The curve of the binomial deviance, defined an
2log likelihood, is plotted versus the log (λ). The dotted vertical lines are drawn at the optimal values by using the minimum criteria and the 1 standard error of the minimum criteria (1-SE criteria). Log (λ) = 2.509 was chosen (i.e., 1-SE criteria).

6

ARTICLE IN PRESS Academic Radiology, Vol &, No &&, && 2019 DIFFERENTIAL DIAGNOSIS OF NONHYPERVASCULAR PANCREATIC

Figure 2 Continued.

radiological features (i.e., MDAC) and texture features (i.e., Maximum3DDiameter). An increased tumor size has been linked to cystic necrosis and the subsequent development of dystrophic calcifications (38). However, PDACs do not calcify, unless they occur in combination with chronic calcific pancreatitis (38,39). This factor explains the large difference in calcification. Compared to PDACs, PNENs are more likely to involve the blood vessels instead of invading the vascular wall and consequently obstructing the lumen, which primarily occurs because of the larger volume and less aggressiveness of PNENs. Jeon et al. (14) also reported that a welldefined tumor margin was a useful MRI feature for the differential diagnosis of nonhypervascular PNENs and PDACs.

Unlike our CT-based study, calcification is less prominent in Jeon’s study, which is due to MRI not being sensitive for presence of calcifications. According to a previous study, texture features such as uniformity represent the homogeneity of tumors’ parenchyma (22). Therefore, the results of the current study, nonhypervascular PNENs tend to have higher uniformity (texture feature) and less heterogeneity (radiological feature), manifest the consistency of these two different analytical approaches to CT images. Three hundred eighty-seven texture features of the tumor were extracted from the CT images. After applying the LASSO process, the following univariate regression could identify features that possess the most powerful correlation 7

ARTICLE IN PRESS Academic Radiology, Vol &, No &&, && 2019

YU ET AL

TABLE 3. Results of the Uni- and Multivariate Logistic Regression Analyses of Features Selected by Using LASSO Texture Feature

AP PVP

Univariate Analysis

Quantile0.025 Maximum3DDiameter ClusterShade_angle135_offset4 HighGrayLevelRunEmphasis_AllDirection_offset7_SD Maximum3DDiameter

Multivariate Analysis

pValue

OR

pValue

OR

0.02 0.00 0.03 0.00 0.00

1.02 (1.00,1.04) 1.04 (1.02,1.06) 1.02 (1.00,1.04) 0.87 (0.79,0.96) 1.04 (1.02,1.06)

0.01 0.00 0.06 0.04 0.00

1.03 (1.01,1.05) 1.04 (1.02,1.06) 1.04 (1.01,1.06) 0.87 (0.75,0.99) 1.05 (1.02,1.08)

The data in the parentheses are the 95% confidence intervals. AP, arterial phase; LASSO, least absolute shrinkage and selection operator; OR, odds ratio; PVP, portal vein phase.

Figure 3. Receiver operating characteristic curve of multivariate logistic regression models for differentiating nonhypervascular PNENs from PDACs PNEN, nonhypervascular pancreatic neuroendocrine neoplasm; PDACs, pancreatic ductal adenocarcinoma; ROC, receiver operating characteristic.

between two types of tumors. Quantile0.025 and Maximum3DDiameter, which belong to the first-order and morphological texture feature, respectively, demonstrate significant differences both in AP images and PVP images. The reproducibility of these two features consists of other studies’ results for first-order and shape CT features were generally more repeatable than other textural features (40
42). However, skewness and uniformity, which are also first-order features, were solely selected in the PVP. We speculate that the reason for such contradictory results is attributable to the

different slice thickness used in the AP and PVP images, which degraded feature reproducibility (43
46). Li (22) reported that skewness, rather than entropy, based on the PVP image, could help differentiate atypical pancreatic neuroendocrine tumors from PDACs. Our results showed some similarity with regard to skewness; however, unlike entropy, GLCM entropy, as a high-order texture feature, shows it has diagnostic potential. The incompatible result may be because of the difference between first-order and high-order texture features. With regard to entropy and uniformity, Guo (47) similarly reported that PDACs have higher entropy and less uniformity than do pancreatic neuroendocrine carcinomas, which are a rare subtype of PNEN, and were proportionately large in our experimental group (9/40), compared to the common PNENs demographic. After applying multivariate logistic regression analysis, multiple features, which were selected with LASSO and univariate logistic regression analysis, did not prove to be independent predictors in the final modeling process, although some of them have been logically justified and verified by the aforementioned studies. This negative result could not negate the usefulness of these features because all selected features were multiplied by different b values and then enrolled into the equation of the model (48). By utilizing the logistic regression analysis method we finally achieved three models based on radiological features, AP texture features, and PVP texture features, respectively. Our results showed that the diagnostic performance of the PVP texture feature model was significantly better than that of the AP texture feature model and the radiological feature model. This finding indicated that in the differential diagnosis of nonhypervascular PNENs and PDACs, the PVP images are more worthwhile than AP images in dual-phasic CT in terms of TA.

TABLE 4. Diagnostic Performance of the Logistic Regression Models Models

Sen (95% CI)

Spe (95% CI)

AUC (95% CI)

Radiological feature model AP texture feature model PVP texture feature model Integrated feature model

65.0 (48.3, 79.4) 80.0 (64.4, 90.9) 85.0 (70.2, 94.3) 97.5 (86.8, 99.9)

85.0 (75.3, 92.0) 73.7 (62.7, 83.0) 88.7 (79.7, 94.7) 66.2 (54.8, 76.4)

0.780 (0.695, 0.850) 0.855 (0.779, 0.913) 0.929 (0.864, 0.966) 0.888 (0.818, 0.938)

AP, arterial phase; AUC, area under the curve; CI, confidence interval; PVP, portal vein phase; Sen, sensitivity; Spe, specificity.

8

ARTICLE IN PRESS Academic Radiology, Vol &, No &&, && 2019 DIFFERENTIAL DIAGNOSIS OF NONHYPERVASCULAR PANCREATIC

Furthermore, compared to the radiological feature model, PVP texture feature models show greater diagnostic efficacy, which demonstrates the promising application of TA. The integrated feature model, which combines every significant predictor from radiological features and texture features, did not show better performance in the differential diagnosis of nonhypervascular PNENs and PDACs. A possible reason is the increased number of variates in the integrated model with a relatively small sample size. Another reason is that some redundant features may exist between AP and PVP texture features, especially those that represent morphological characteristics of the tumor. Our study had some limitations. First, the study was retrospective and the number of patients was relatively small. Second, the section thickness of AP images was uniform (5 mm) in this study, although the different section thicknesses of the PVP images and different CT machines may have caused variations in the texture features. Third, we had only one TA reviewer in this study, although previous TA studies have shown good to excellent interobserver agreement (49,50).

CONCLUSION Our study results suggest that a well-defined margin, calcification, homogeneity, and blood vessels in the tumor were useful radiological features for discriminating nonhypervascular PNENs from PDACs. TA of contrast-enhanced CT can be a beneficial tool in the differential diagnosis of nonhypervascular PNENs and PDACs. In future studies, larger dataset of patients with standardized scanning protocol is expected, which would make our model building approach more sophisticated more reproducible. REFERENCES 1.

Rindi G, Wiedenmann B. Neuroendocrine neoplasms of the gut and pancreas: new insights. Nat Rev Endocrinol. 2011; 8(1):54–64. 2. Lewis RB, Lattin Jr. GE, Paal E. Pancreatic endocrine tumors: radiologicclinicopathologic correlation. Radiographics 2010; 30(6):1445–1464. 3. Lee NJ, Hruban RH, Fishman EK. Pancreatic neuroendocrine tumor: review of heterogeneous spectrum of CT appearance. Abdom Radiol (NY) 2018; 43(11):3025–3034. 4. Modlin IM, Oberg K, Chung DC, et al. Gastroenteropancreatic neuroendocrine tumours. Lancet Oncol 2008; 9(1):61–72. 5. Zerbi A, Falconi M, Rindi G, et al. Clinicopathological features of pancreatic endocrine tumors: a prospective multicenter study in Italy of 297 sporadic cases. Am J Gastroenterol 2010; 105(6):1421–1429. 6. Yao JC, Hassan M, Phan A, et al. One hundred years after “carcinoid”: epidemiology of and prognostic factors for neuroendocrine tumors in 35,825 cases in the United States. J Clin Oncol 2008; 26(18):3063–3072. 7. National Comprehensive Cancer Network. (NCCN) Clinical Practice Guidelines in Oncology Pancreatic adenocarcinoma. July10, 2018 [cited 2018 Sept 1]; Version 2. 2018: [https://www.nccn.org/professionals/ physician_gls/default.aspx#site]. 8. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA Cancer J Clin 2019; 69(1):7–34. 9. Strosberg JR, Cheema A, Weber J, Han G, Coppola D, Kvols LK. Prognostic validity of a novel American Joint Committee on Cancer Staging Classification for pancreatic neuroendocrine tumors. J Clin Oncol 2011; 29(22):3044–3049. 10. Adamska A, Domenichini A, Falasca M. Pancreatic ductal adenocarcinoma: current and evolving therapies. Int J Mol Sci 2017; 18(7).

11. Vincent A, Herman J, Schulick R, Hruban RH, Goggins M. Pancreatic cancer. Lancet (London, England) 2011; 378(9791):607–620. 12. Ichikawa T, Peterson MS, Federle MP, et al. Islet cell tumor of the pancreas: biphasic CT versus MR imaging in tumor detection. Radiology 2000; 216(1):163–171. 13. Stafford-Johnson DB, Francis IR, Eckhauser FE, Knol JA, Chang AE. Dual-phase helical CT of nonfunctioning islet cell tumors. J Comput Assist Tomogr 1998; 22(2):335–339. 14. Jeon SK, Lee JM, Joo I, et al. Nonhypervascular pancreatic neuroendocrine tumors: differential diagnosis from pancreatic ductal adenocarcinomas at MR imaging-retrospective cross-sectional study. Radiology 2017; 284(1):77–87. 15. De Robertis R, Tinazzi Martini P, Cingarlini S, et al. Digital subtraction of magnetic resonance images improves detection and characterization of pancreatic neuroendocrine neoplasms. J Comput Assist Tomogr 2017; 41(4):614–618. 16. Belousova E, Karmazanovsky G, Kriger A, et al. Contrast-enhanced MDCT in patients with pancreatic neuroendocrine tumours: correlation with histological findings and diagnostic performance in differentiation between tumour grades. Clin Radiol 2017; 72(2):150–158. 17. Pietryga JA, Morgan DE. Imaging preoperatively for pancreatic adenocarcinoma. J Gastrointest Oncol 2015; 6(4):343–357. 18. Gallotti A, Johnston RP, Bonaffini PA, et al. Incidental neuroendocrine tumors of the pancreas: MDCT findings and features of malignancy. AJR Am J Roentgenol. 2013; 200(2):355–362. 19. Horton KM, Hruban RH, Yeo C, Fishman EK. Multi-detector row CT of pancreatic islet cell tumors. Radiographics 2006; 26(2):453–464. 20. Guillen-Ponce C, Blazquez J, Gonzalez I, de-Madaria E, Montans J, Carrato A. Diagnosis and staging of pancreatic ductal adenocarcinoma. Clin Transl Oncol 2017; 19(10):1205–1216. 21. Parbhu SK, Adler DG. Pancreatic neuroendocrine tumors: contemporary diagnosis and management. Hospital Practice (1995) 2016; 44(3):109–119. 22. Li J, Lu J, Liang P, et al. Differentiation of atypical pancreatic neuroendocrine tumors from pancreatic ductal adenocarcinomas: using wholetumor CT texture analysis as quantitative biomarkers. Cancer Med 2018; 7(10):4924–4931. 23. Galloway MM. Texture analysis using grey level run lengths. Computer Graphics and Image Processing 1975; 4(2):172–179. 24. Ganeshan B, Panayiotou E, Burnand K, Dizdarevic S, Miles K. Tumour heterogeneity in non-small cell lung carcinoma assessed by CT texture analysis: a potential marker of survival. Eur Radiol 2012; 22(4):796–802. 25. Miles KA, Ganeshan B, Rodriguez-Justo M, et al. Multifunctional imaging signature for V-KI-RAS2 Kirsten rat sarcoma viral oncogene homolog (KRAS) mutations in colorectal cancer. J Nucl Med. 2014; 55(3):386–391. 26. Zhang H, Graham CM, Elci O, et al. Locally advanced squamous cell carcinoma of the head and neck: CT texture and histogram analysis allow independent prediction of overall survival in patients treated with induction chemotherapy. Radiology 2013; 269(3):801–809. 27. Kim JH, Ko ES, Lim Y, et al. Breast cancer heterogeneity: MR imaging texture analysis and survival outcomes. Radiology 2017; 282(3):665–675. 28. Sandrasegaran K, Lin Y, Asare-Sawiri M, Taiyini T, Tann M. CT texture analysis of pancreatic cancer. Eur Radiol 2019; 29(3):1067–1073. 29. Hanania AN, Bantis LE, Feng Z, et al. Quantitative imaging to evaluate malignant potential of IPMNs. Oncotarget 2016; 7(52):85776–85784. 30. Attiyeh MA, Chakraborty J, Gazit L, et al. Preoperative risk prediction for intraductal papillary mucinous neoplasms by quantitative CT image analysis. HPB (Oxford) 2018. 31. Huang Z, Li M, He D, et al. Two-dimensional texture analysis based on CT images to differentiate pancreatic lymphoma and pancreatic adenocarcinoma: a preliminary study. Acad Radiol 2019; 26(8):e189–e195. 32. Galloway MM. Texture analysis using grey level run lengths. NASA STI/ Recon Technical Report N 1974; 75. 33. Thibault G, Fertil B, Navarro C, et al. Texture indexes and gray level size zone matrix. Application to cell nuclei classification. In: Pattern Recognition and Information Processing (PRIP); 2009. p. 140–145. 34. Tibshirani Robert. Regression shrinkage and selection via the lasso. J R Stat Soc Ser B (Methodological) 1996; 58(1):267–288. 35. Vasquez MM, Hu C, Roe DJ, Chen Z, Halonen M, Guerra S. Least absolute shrinkage and selection operator type methods for the identification of serum biomarkers of overweight and obesity: simulation and application. BMC Med Res Method 2016; 16(1):154. 36. Rodallec M, Vilgrain V, Couvelard A, et al. Endocrine pancreatic tumours and helical CT: contrast enhancement is correlated with microvascular

9

ARTICLE IN PRESS YU ET AL

37.

38. 39. 40.

41.

42.

43.

10

density, histoprognostic factors and survival. Pancreatology 2006; 6(1-2):77–85. Legrand L, Duchatelle V, Molinie V, Boulay-Coletta I, Sibileau E, Zins M. Pancreatic adenocarcinoma: MRI conspicuity and pathologic correlations. Abdom Imaging 2015; 40(1):85–94. Verde F, Fishman EK. Calcified pancreatic and peripancreatic neoplasms: spectrum of pathologies. Abdom Radiol (NY) 2017; 42(11):2686–2697. Lesniak RJ, Hohenwalter MD, Taylor AJ. Spectrum of causes of pancreatic calcifications. AJR Am J Roentgenol. 2002; 178(1):79–86. Desseroit M-C, Visvikis D, Tixier F, et al. Development of a nomogram combining clinical staging with 18F-FDG PET/CT image features in non-small-cell lung cancer stage I
III. Eur J Nucl Med Mol Imaging 2016; 43(8):1477–1485. Tixier F, Hatt M, Le Rest CC, Le Pogam A, Corcos L, Visvikis D. Reproducibility of tumor uptake heterogeneity characterization through textural feature analysis in 18F-FDG PET. J Nucl Med. 2012; 53(5):693–700. van Timmeren JE, Leijenaar RTH, van Elmpt W, et al. Test-retest data for radiomics feature stability analysis: generalizable or study-specific? Tomography (Ann Arbor, Mich) 2016; 2(4):361–365. Fave X, Zhang L, Yang J, et al. Impact of image preprocessing on the volume dependence and prognostic potential of radiomics features in nonsmall cell lung cancer. Transl Cancer Res 2016; 5(4):349–363.

Academic Radiology, Vol &, No &&, && 2019

44. Bagher-Ebadian H, Siddiqui F, Liu C, Movsas B, Chetty IJ. On the impact of smoothing and noise on robustness of CT and CBCT radiomics features for patients with head and neck cancers. Med Phys 2017; 44 (5):1755–1770. 45. Zhao B, Tan Y, Tsai WY, et al. Reproducibility of radiomics for deciphering tumor phenotype with imaging. Sci Rep 2016; 6:23428. 46. Bogowicz M, Riesterer O, Bundschuh RA, et al. Stability of radiomic features in CT perfusion maps. Phys Med Biol 2016; 61(24):8736–8749. 47. Guo C, Zhuge X, Wang Q, et al. The differentiation of pancreatic neuroendocrine carcinoma from pancreatic ductal adenocarcinoma: the values of CT imaging features and texture analysis. Cancer Imaging 2018; 18(1):37. 48. Sperandei S. Understanding logistic regression analysis. Biochem Med 2014; 24(1):12–18. 49. Hodgdon T, McInnes MD, Schieda N, Flood TA, Lamb L, Thornhill RE. Can quantitative CT texture analysis be used to differentiate fat-poor renal angiomyolipoma from renal cell carcinoma on unenhanced CT images? Radiology 2015; 276(3):787–796. 50. Smith AD, Gray MR, del Campo SM, et al. Predicting overall survival in patients with metastatic melanoma on antiangiogenic therapy and RECIST stable disease on initial posttherapy images using CT texture analysis. AJR Am J Roentgenol. 2015; 205(3):W283–W293.