Applying a radiomics-based strategy to preoperatively predict lymph node metastasis in the resectable pancreatic ductal adenocarcinoma

Abstract

PURPOSE:

This retrospective study is designed to develop a Radiomics-based strategy for preoperatively predicting lymph node (LN) status in the resectable pancreatic ductal adenocarcinoma (PDAC) patients.

METHODS:

Eighty-five patients with histopathological confirmed PDAC are included, of which 35 are LN metastasis positive and 50 are LN metastasis negative. Initially, 1,124 radiomics features are computed from CT images of each patient. After a series of feature selection, a Radiomics logistic regression (LOG) model is developed. Subsequently, the predictive efficiency of the model is validated using a leave-one-out cross-validation method. The model performance is evaluated on discrimination and compared with the conventional CT evaluation method based on subjective CT image features.

RESULTS:

Radiomics LOG model is developed based on eight most related radiomics features. Remarkable differences are demonstrated between patients with LN metastasis positive and LN metastasis negative in Radiomics LOG scores namely, 0.535±1.307 (mean±standard deviation) vs. −1.514±1.800 (mean±standard deviation) with p < 0.001. Radiomics LOG model shows significantly higher predictive efficiency compared to the conventional evaluation method of LN status in which areas under ROC curves are AUC = 0.841 with 95% confidence interval (CI: 0.758∼0.925) vs. AUC = 0.682 with (95% CI: 0.566∼0.798). Leave-one-out cross validation indicates that the Radiomics LOG model correctly classifies 70.3% cases, while the conventional CT evaluation method only correctly classifies 57.0% cases.

CONCLUSION:

A radiomics-based strategy provides an individualized LN status evaluation in PDAC patients, which may help clinicians implement an optimal personalized patient treatment.

Keywords

Computed tomography radiomics pancreatic ductal Adenocarcinoma lymph node metastasis personalized medicine

1 Introduction

Pancreatic cancer is the deadliest malignancy, and pancreatic ductal adenocarcinoma (PDAC) is the main type of pancreatic cancer [1, 2]. Over the past few decades, many types of cancers showed a stable or decreasing trends of incidence and mortality rates. However, the incidence of PDAC continues to rise and the prognosis is still poor [3, 4]. Radical pancreatic resection is the only potentially curative therapy for PDAC. However, the patients who suitable for radical resection only account for 10–15%, and the majority of patients are un-resectable or already have metastatic disease [5]. Lymph node status is one of the factors affecting whether to undergo radical treatment. Previous studies showed that lymph node metastases can be found in 70–80% of surgically resected patients with PDAC [6 –9] and is strongly correlated with poor prognosis [10]. The Evaluation of lymph node status has been regarded as a significant factor when determining personalized therapy method for cancer patients [11 –14].

Surgical staging at the time of pancreatic resection is the gold standard for the accurate N staging of PDAC. However, it is invasive and can only be obtained after surgery. Currently, computed tomography (CT) is the most common imaging modality for PDAC therapy guidance and tumor monitor. However, previous studies have shown that examination of peri-pancreatic lymph nodes by conventional CT cannot predict the presence of metastatic deposits [15]. Other noninvasive strategies include MRI and positron emission tomography (PET), but the previous studies have also showed that none of them can accurately preoperative identification of LN metastases [16].

Radiomics is an emerging field focused on extracting quantitative information from medical images through advanced imaging processing and analysis tools [17]. Research studies have demonstrated that radiomics can be used to evaluate LN metastases in bladder cancer, biliary tract cancer and Lung cancer [18 –20]. However, few studies have used radiomics to predict PDAC LN metastases. Therefore, this study intends to develop and validate a predictive model based on the radiomics strategy to predict PDAC LN metastases non-invasively, enabling a better personalized evaluation and treatment.

2 Materials and methods

2.1 Patient data

In total, 85 patients underwent curative intent surgical resection for PDAC were collected. The inclusion criteria were: (1) Histologically confirmed diagnosis of PDAC with primary tumor resection and standard lymphadenectomy [21]; (2) availability of contrast-enhanced CT images for analysis, 2 weeks before surgery; (3) not received any treatment (chemotherapy, radiotherapy, or chemoradiotherapy) before their CT scan. The exclusion criteria were: (1) patients with other concurrent primary malignant tumors (2) those CT images with an unclear pancreatic mass. (3) incomplete clinical data. Finally, a total of 85 patients were included, of which 35 patients were LN metastasis positive (LN+) and 50 patients were LN metastasis negative (LN−). Our institutional ethics board approved this retrospective study and waived the requirement for informed consent.

2.2 Image acquisitions

The CT examinations were performed using quadruple-phase on Philips iCT 256 (Philips, Amsterdam, Netherlands), which consisted of unenhanced, arterial phases, venous phases and equilibrium phases. The scanning parameters were as follows: the tube voltage 120 kV, tube current modulation 200–380 mA, field of view 380–500mm, and reconstructed section thickness 5 mm. About per 1.2–1.5 mL of iodinated contrast (Omnipaque 300, GE Healthcare) per kilogram of body weight was administered using a power injector at 3.0 mL/s after unenhanced scan. Arterial phase images were obtained 25–35s after the trigger threshold, 60–70s for the venous phase and 3min for the equilibrium phase.

2.3 CT subjective image features evaluation and establishment of conventional CT evaluation model

According to the related literature [22 –24], we evaluated several conventional CT features which may be useful for predicting LN metastasis. These features including primary tumor location, pancreatic duct dilatation, bile duct dilatation, ovoid LN, non-uniform density LN and LN short diameter. LN short diameter was measured on the trans-axial image of CT, an LN with a short diameter larger than 1 cm in the short axis was predicted as positive (LN+); otherwise, it was predicted as negative (LN−). All these measurements were done by two experienced radiologists (with 5 and 9 years of clinical experience in abdomen CT interpretation, respectively, and blinded to patient LN metastasis status). Their disagreement on measurement was resolved by consensus. The positive result was recorded as 1, the negative result was recorded as o. Finally, all these features were run through a logistic regression (LOG) and a conventional LOG model was built.

2.4 ROI segmentation and Radiomics features extraction

This study used the manual segmentation function of 3D-slicer (Version 4.11.0, https://www.slicer.org/) to segment the region of interest (ROI). To evaluate the inter-observer reproducibility, 2 radiologists performed the ROI segmentation in a blinded fashion, respectively. Both knew the PDAC diagnosis but did not know the clinical and LN metastasis status. To assess the intra-observer reproducibility, reader 1 repeated the feature extraction twice in a 2-week period. Both two readers were aware of contouring the ROI on the slice with the largest visible cross section of the tumor on the venous phase (the phase during which pancreatic cancer boundaries were most consistently seen). The ROI was within the tumor and close to the tumor margin [25].

The radiomics features were automatically extracted from the ROIs by using the SlicerRadiomics (an extension in 3D-slicer). which include 93 original texture features, 8 morphology features, and 1023 high-order features (279 Gaussian filter transform features and 744 Wavelet transform features). In total, 1124 radiomic features were extracted in this study. A detailed description of the radiomics features can be found in previous studies [26, 27]. The reproducibility of radiomics features in inter/intra-observer was reflected according to the intra-class correlation coefficients (ICCs) and we chose 50 random ROI segmentations for calculating ICCs.

2.5 Statistics analysis

The workflow of radiomics analysis is summarized in Fig. 1. All analyses were performed on SPSS 20.0 (IBM, Armonk, NY, USA) or R software (V3.5.0, https://www.R-project.org/). Independent sample t-test or χ2 test was used to assess the differences of clinical characteristics between LN+ and LN− group. The following four steps were implemented to identify significant radiomics features. First, the Mann-Whitney U-test was applied to select radiomics features that significantly different between LN+ and LN− group (p < 0.05). Second, we excluded features with ICC <0.3 to eliminate unstable features [28]. Third, we filtered features based on their independence from others by using spearman correlation test. Finally, feature selection was done by using the least absolute shrinkage and selection operator (LASSO) method. The remaining non-zero coefficient features were run through a logistic regression (LOG) model (“glmnet” package in R software). Then, performance was validated by a leave-one-out cross-validation method. Receiver operating characteristic (ROC) curve was used to evaluate the diagnostic performance of LOG model. Corresponding areas under the curve (AUCs) were calculated. The DeLong test was used to ROC comparison between LOG model and conventional evaluation method. All with P value <0.05 considered as statistically significant.

Fig. 1

Radiomics analysis workflow of this study.

3 Result

3.1 Clinical features

Table 1 listed the clinical characteristics of patients. In total, 85 patients were included. (median age: 60 years, age range: 37–80 years) In univariate analyses, the results showed that there were no significant differences between LN+ and LN− groups in clinical characteristics (P > 0.05).

Table 1
Characteristics of the study population

Characteristic Total (n = 85) LN+ (n = 35) LN− (n = 50) P

Gender

Male 46 (54.1%) 19 (54.3%) 27 (54.0%) 0.979

Female 39 (45.9%) 16 (45.7%) 23 (46.0%)

Age (years)

Mean (±SD) 59.4 (±9.4) 60.4 (±8.6) 58.6 (±10.0) 0.392

Primary tumor location

Body or tail 15 (17.6%) 3 (8.6%) 12 (24.0%) 0.066

Head 70 (82.4%) 32 (91.4%) 38 (76.0%)

PDD

No 13 (15.3%) 5 (14.3%) 8 (16.0%) 0.829

Yes 72 (84.7%) 30 (85.7%) 42 (84.0%)

BD

No 32(37.6%) 14 (40.0%) 18 (36.0%) 0.708

Yes 53 (62.4%) 21 (60.0%) 32 (64.0%)

Ovoid LN

No 66 (77.6%) 27 (77.1%) 39 (78.0%) 0.926

Yes 19 (22.4%) 8(22.9%) 11 (21.0%)

Nonuniform density LN

No 80 (94.1%) 34 (97.1%) 46 (92.0%) 0.601

Yes 5 (5.9%) 1 (2.9%) 4 (8.0%)

CT-based LN status [n(%)]

LN+ 10(11.8%) 5(14.3%) 5(10.0%) 0.794

LN− 75(88.2%) 30(85.7%) 45(90.0%)

Characteristic	Total (n = 85)	LN+ (n = 35)	LN− (n = 50)	P
Gender
Male	46 (54.1%)	19 (54.3%)	27 (54.0%)	0.979
Female	39 (45.9%)	16 (45.7%)	23 (46.0%)
Age (years)
Mean (±SD)	59.4 (±9.4)	60.4 (±8.6)	58.6 (±10.0)	0.392
Primary tumor location
Body or tail	15 (17.6%)	3 (8.6%)	12 (24.0%)	0.066
Head	70 (82.4%)	32 (91.4%)	38 (76.0%)
PDD
No	13 (15.3%)	5 (14.3%)	8 (16.0%)	0.829
Yes	72 (84.7%)	30 (85.7%)	42 (84.0%)
BD
No	32(37.6%)	14 (40.0%)	18 (36.0%)	0.708
Yes	53 (62.4%)	21 (60.0%)	32 (64.0%)
Ovoid LN
No	66 (77.6%)	27 (77.1%)	39 (78.0%)	0.926
Yes	19 (22.4%)	8(22.9%)	11 (21.0%)
Nonuniform density LN
No	80 (94.1%)	34 (97.1%)	46 (92.0%)	0.601
Yes	5 (5.9%)	1 (2.9%)	4 (8.0%)
CT-based LN status [n(%)]
LN+	10(11.8%)	5(14.3%)	5(10.0%)	0.794
LN−	75(88.2%)	30(85.7%)	45(90.0%)

LN, lymph nodes; LN+, lymph nodes metastasis positive; LN−, lymph nodes metastasis negative; PDD, pancreatic duct dilatation; BD, bile duct dilatation; CT, computed tomography; SD, standard deviation.

3.2 Performances of conventional LOG model

In univariate analyses, the result showed that no conventional features could be used to discriminate the LN+ group and LN− group with an AUC ranging from 0.506 to 0.552. Putting these features into the logistic regression analysis, the conventional LOG model was constructed and the calculation formula for the conventional LOG model scores was (1). The results showed that the AUC value of conventional LOG model in predicting LN status was 0.682, (95% CI, 0.566∼0.798), with sensitivity and specificity of 0.486 and 0.780. Leave-one-out cross validation indicated the conventional LOG model could classify 57.0% cases accurately (Fig. 2).

Fig. 2

The comparison of Conventional logistic regression (LOG) model and Radiomics logistic regression (LOG) model. (a) ROC curve of the Conventional LOG model, (b) ROC curve of the Radiomics LOG model, (c) Radiomcs LOG model scores for each patient, (d) Conventional LOG model scores for each patient.

$\begin{matrix} Conventional LOG model scores \\ = 2.0332 \times PrimaryTumorlocation - 0.7720 \times Pancreaticductdilatation - 0.9099 \\ \times Bileductdilatation - 0.2695 \times OvoidLN + 1.1024 \times LNshortdiameter \\ - 1.3686 \times NonuniformdensityLN - 0.8490 \end{matrix}$ (1)

3.3 Feature selection and establishment of radiomics LOG model

Among the 1124 radiomics features, 14 features were retained through the Mann-Whitney U-test. Feature stability was evaluated by using the ICC. A total of 12 features with ICC >0.3 were retained; the result of ICC test is shown in Table 2. After that, 1 feature was excluded because of significant correlation with another by Spearman test. Then, 11 radiomics features were used for LASSO feature selection. Finally, 8 radiomics features with non-zero coefficients in LASSO analysis were selected and the Radiomics LOG model was subsequently constructed based on the logistic regression. The calculation formula for the Radiomics LOG model scores was (2).

Table 2
Evaluation for intra-observer agreement of the extracted radiomics feature by intra-class correlation coefficients (ICCs)

Raiomics features intra-class inter-class

InterquartileRange 0.316 0.560

RunLengthNonUniformityNormalized 0.713 0.694

Median 0.419 0.911

SmallDependenceLowGrayLevelEmphasis 0.650 0.508

SmallAreaEmphasis 0.370 0.490

SizeZoneNonUniformityNormalized 0.723 0.467

SizeZoneNonUniformity 0.676 0.722

SmallDependenceEmphasis 0.722 0.760

SizeZoneNonUniformityNormalizedII 0.592 0.549

Kurtosis 0.773 0.470

JointEnergy 0.938 0.554

MaximumProbability 0.926 0.674

Raiomics features	intra-class	inter-class
InterquartileRange	0.316	0.560
RunLengthNonUniformityNormalized	0.713	0.694
Median	0.419	0.911
SmallDependenceLowGrayLevelEmphasis	0.650	0.508
SmallAreaEmphasis	0.370	0.490
SizeZoneNonUniformityNormalized	0.723	0.467
SizeZoneNonUniformity	0.676	0.722
SmallDependenceEmphasis	0.722	0.760
SizeZoneNonUniformityNormalizedII	0.592	0.549
Kurtosis	0.773	0.470
JointEnergy	0.938	0.554
MaximumProbability	0.926	0.674

$\begin{matrix} Radiomics LOG model scores \\ = 0.04107 \times Log - Sigma - 6 - 0 - mm - 3 D_firstorder_Median - 23.41570 \\ \times Log - sigma - 6 - 0 mm - 3 D_gldm_SmallDependenceLowGrayLevelEmphasis - 2.75887 \\ \times Log - sigma - 6 - 0 mm - 3 D_glszm_SmallAreaEmphasis - 2.38928 \times Log - sigma - 2 - 0 mm \\ - 3 D_glszm_SizeZoneNonUniformityNormalize + 0.14204 \times Wavelet_LLH_glszm_Siz \\ eZoneNonUniformity - 6.04189 \times Wavelet - LHL_glszm_SizeZoneNonUniformityNormalize d \\ - 0.86199 \times Original_firstorder_Kurtosis - 0.64700 \times Original_glcm_MaximumProba bilit \\ + 5.78020 \end{matrix}$ (2)

3.4 Performance of Radiomics LOG model

Significant differences were found between patients with LN+ and LN− in Radiomics LOG model scores (0.535±1.307 (mean±standard deviation) vs. −1.514±1.800 (mean±standard deviation), p < 0.001). The optimal threshold of 0.439 was selected by using the Youden Index in the ROC analysis. Patients with the Radiomics LOG model scores greater than 0.439 were predicted as LN+, while patients with the scores lower than 0.439 were predicted as LN−. The AUC value was 0.841, (95% CI, 0.758∼0.925), with sensitivity and specificity of 0.800 and 0.780. 70.3% cases were correctly classified using leave-one-out cross-validation (Fig. 2).

Delong test was performed on the ROC curves of Radiomics LOG model and conventional LOG model. The results showed that there was a statistically significant difference between these two models (AUC, 0.841 vs. 06825; sensitivity, 0.800 vs. 0.486; specificity, 0.780 vs. 0.780; accuracy, 70.3% vs. 57.0%). It showed that Radiomics LOG model can significantly improve the prediction accuracy.

4 Discussion

In this study, we sought to identify a new method to predict LN metastasis in patients with PDAC based on pathologically confirmed results. Finally, a radiomics-based model was built and achieved a good performance.

Precise identification of pancreatic cancer LN metastasis before surgery can help guide individualized diagnosis and treatment of pancreatic cancer. The gold standard for LN status evaluation is pathology. However, it is invasive and expensive. CT is the most used method for LN evaluation and an LN diameter of more than 10 mm is considered as LN+. However, it has a very low accuracy, especially in poor sensitivity (20% –38%) [16 , 30]. The other criteria for metastatic LN such as nonuniform density, shape, BD, PD et al. also yield low diagnostic accuracy and sensitivity. In concordance with prior studies, we found that there were no conventional CT features could be used to discriminate the LN+ group and LN− group. When all those conventional features were put into model analysis, the result was still poor (AUC, 0.682; sensitivity, 0.486; specificity, 0.780, accuracy, 57%). This result reflected the limitation ability of conventional CT features to predict LN status.

Thus, we attempted to adopt a radiomics strategy to predict LN metastasis in patients with PADC. The important implication of our results is that the radiomics-based strategy (Radiomics LOG model) showed great performance in predicting the statue of LN metastasis in patients with PDAC. Radiomics is a novel medical imaging analysis technology, which can yield tissue-specific information using radiomics features that analyze the change in image pixels and grey-scale spatial distribution [31]. A previous study by Bian Y et al. [32]. explored the relationship between the radiomics score and LN metastasis in PDAC. In contrast to the study by Bian Y et al., our study focused on developing a prediction model. The model showed good discrimination and validated by a leave-one-out cross-validation method.

The strength of this study is that the Radiomics LOG model was superior to conventional LOG model. In the previous studies, various modalities including CT, PET/CT, and MRI were used to distinguish malignant and benign LNs. However, all of them were trying to directly image metastatic lymph nodes. Actually, because of resolution, it is hard to detect LN. Studies showed that there is high lymphatic vessel density (LVD) in PDAC tumors or peri-tumors [33, 34]. Many studies also agree that PDAC invades lymph nodes frequently through the lymphatic vasculature in the tumour rather than through direct/contiguous extension of the primary tumor to the lymph node [35, 36]. So, we presume that predicting LN metastasis by analyzing primary tumor maybe better than imaging LN. Many radiomics-based strategies had been successful in predicting LN metastasis by directly analyzing primary tumor.

The present study also has its limitations, First, we included a relatively small number of patients and lacked an external validation. Multicenter validation with a larger sample size is needed before universal application of these results. In addition, adding more clinical information and multi-phase CT information may improve the performance of the prediction model.

5 Conclusion

In conclusion, this study showed that a radiomics-based strategy may provide a novel non-invasive strategy that preoperatively predicts LN metastasis in PDAC patients.

Conflict of interest statement

The authors have no conflict of interest to report.

References

Leinonen

M.K.

, Miettinen

, Heikkinen

, et al., Quality measures of the population-based Finnish Cancer Registry indicate sound data quality for solid malignant tumours, Eur J Cancer 77 (2017), 31–39.

Dimitrakopoulos

, Vrugt

, Flury

, et al., Identification and validation of a biomarker signature in patients with resectable pancreatic cancer via genome-wide screening for functional genetic variants, JAMA Surg 154(6) (2019), e190484.

Siegel

R.L.

, Miller

K.D.

and Jemal

, Cancer statistics, 2019, CA Cancer J Clin 69(1) (2019), 7–34.

Garrido-Laguna

and Hidalgo

, Pancreatic cancer: from state-of-the-art treatments to promising novel therapies, Nat Rev Clin Oncol 12(6) (2015), 319–334.

Ethun

C.G.

and Kooby

D.A.

, The importance of surgical margins in pancreatic cancer, J Surg Oncol 113(3) (2016), 283–288.

Katz

M.H.

, Hwang

, Fleming

J.B.

, et al., Tumor-node-metastasis staging of pancreatic adenocarcinoma, CA Cancer J Clin 58(2) (2008), 111–125.

Massucco

, Ribero

, Sgotto

, et al., Prognostic significance of lymph node metastases in pancreatic head cancer treated with extended lymphadenectomy: not just a matter of numbers, Ann Surg Oncol 16(12) (2009), 3323–3332.

Murakami

, Uemura

, Sudo

, et al., Number of metastatic lymph nodes, but not lymph node ratio, is an independent prognostic factor after resection of pancreatic carcinoma, J Am Coll Surg 211(2) (2010), 196–204.

Pawlik

T.M.

, Gleisner

A.L.

, Cameron

J.L.

, et al., Prognostic relevance of lymph node ratio following pancreaticoduodenectomy for pancreatic cancer, Surgery 141(5) (2007), 610–618.

10.

Robinson

S.M.

, Rahman

, Haugk

, et al., Metastatic lymph node ratio as an important prognostic factor in pancreatic ductal adenocarcinoma, Eur J Surg Oncol 38(4) (2012), 333–339.

11.

Liu

, Luo

, Guo

, et al., Lymph node status predicts the benefit of adjuvant chemoradiotherapy for patients with resected pancreatic cancer, Pancreatology 15(3) (2015), 253–258.

12.

Nathanson

S.D.

, Shah

and Rosso

, Sentinel lymph node metastases in cancer: causes, detection and their role in disease progression, Semin Cell Dev Biol 38 (2015), 106–116.

13.

Cao

HST

, Zhang

, Sada

Y.H.

, et al., Value of lymph node positivity in treatment planning for early stage pancreatic cancer, Surgery 162(3) (2017), 557–567.

14.

Stacker

S.A.

, Williams

S.P.

, Karnezis

, et al., Lymphangiogenesis and lymphatic vessel remodelling in cancer, Nature Reviews Cancer 14(3) (2014), 159–172.

15.

Roche

C.J.

, Hughes

M.L.

, Garvey

C.J.

, et al., CT and pathologic assessment of prospective nodal staging in patients with ductal adenocarcinoma of the head of the pancreas, AJR Am J Roentgenol 180(2) (2003), 475–480.

16.

Imai

, Doi

, Kanazawa

, et al., Preoperative assessment of para-aortic lymph node metastasis in patients with pancreatic cancer, Int J Clin Oncol 15(3) (2010), 294–300.

17.

Aerts

H.J.

, Velazquez

E.R.

, Leijenaar

R.T.

, et al., Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach, Nat Commun 5 (2014), 4006.

18.

, Zheng

, Li

, et al., A radiomics nomogram for the preoperative prediction of lymph node metastasis in bladder cancer, Clin Cancer Res 23(1) (2017), 6904–6911.

19.

Laghi

and Voena

, CT-based radiomics for biliary tract cancer: A possible solution for predicting lymph node metastases, Radiology 290(1) (2018), 99–100.

20.

Sun

, Hu

, Ge

, et al., Radiomics study for predicting the expression of PD-L1 in non-small cell lung cancer based on CT images and clinicopathologic features, J Xray Sci Technol 28(3) (2020), 449–459.

21.

Tol

J.A.

, Gouma

D.J.

, Bassi

, et al., Definition of a standard lymphadenectomy in surgery for pancreatic ductal adenocarcinoma: A consensus statement by the International Study Group on Pancreatic Surgery (ISGPS), Surgery 156(3) (2014), 591–600.

22.

Tseng

D.S.

, Van Santvoort

H.C.

, Fegrachi

, et al., Diagnostic accuracy of CT in assessing extra-regional lymphadenopathy in pancreatic and peri-ampullary cancer: a systematic review and meta-analysis, Surgical Oncology 23(4) (2014), 229–235.

23.

Tanaka

, Heckler

, Mihaljevic

A.L.

, et al., CT response of primary tumor and CA19-9 predict resectability of metastasized pancreatic cancer after FOLFIRINOX,0, Eur J Surg Oncol 45(8) (2019), 1453–1459.

24.

Hwang

S.H.

, Kim

H.Y.

, Lee

E.J.

, et al., Preoperative clinical and computed tomography (CT)-based nomogram to predict oncologic outcomes in patients with pancreatic head cancer resected with curative intent: a retrospective study, J Clin Med 8(10) (2019), 1749.

25.

Sandrasegaran

, Lin

, Asare-Sawiri

, et al., CT texture analysis of pancreatic cancer, Eur Radiol 29(3) (2019), 1067–1073.

26.

Wang

, Gong

, Duan

, et al., Correlation between CT based radiomics features and gene expression data in non-small cell lung cancer, J Xray Sci Technol 27(5) (2019), 773–803.

27.

, Zhu

, Qiu

, et al., Prediction of immunohistochemistry of suspected thyroid nodules by use of machine learning-based radiomics, AJR Am J Roentgenol 213(6) (2019), 1348–1357.

28.

Khalvati

, Zhang

, Baig

, et al., Prognostic value of CT radiomic features in resectable pancreatic ductal adenocarcinoma, Sci Rep 9(1) (2019), 5449.

29.

Masuda

, Dann

A.M.

, Elliott

I.A.

, et al., A Comprehensive assessment of accurate lymph node staging and preoperative detection in resected pancreatic cancer, J Gastrointest Surg 22(2) (2018), 295–302.

30.

Nanashima

, Tobinaga

, Abo

, et al., Evaluation of surgical resection for pancreatic carcinoma at a Japanese single cancer institute, Hepatogastroenterology 59(115) (2012), 911–915.

31.

Limkin

E.J.

, Sun

, Dercle

, et al., Promises and challenges for the implementation of computational medical imaging (radiomics) in oncology, Ann Oncol 28(6) (2017), 1191–1206.

32.

Bian

, Guo

, Jiang

, et al., Relationship between radiomics and risk of lymph node metastasis in pancreatic ductal adenocarcinoma,;, Pancreas 48(9) (2019), 1195–1203.

33.

Wang

, Wu

, Li

, et al., Lymphangiogenesis and biological behavior in pancreatic carcinoma and other pancreatic tumors, Mol Med Rep 5(4) (2012), 959–963.

34.

Kurahara

, Takao

, Shinchi

, et al., Significance of lymphangiogenesis in primary tumor and draining lymph nodes during lymphatic metastasis of pancreatic head cancer, J Surg Oncol 102(7) (2010), 809–815.

35.

Buc

, Couvelard

, Kwiatkowski

, et al., Adenocarcinoma of the pancreas: does prognosis depend on mode of lymph node invasion? Eur J Surg Oncol 40(11) (2014), 1578–1585.

36.

Pai

R.K.

, Beck

A.H.

, Mitchem

, et al., Pattern of lymph node involvement and prognosis in pancreatic adenocarcinoma: direct lymph node invasion has similar survival to node-negative disease, Am J Surg Pathol 35(2) (2011), 228–234.