The multicenter retrospective study was derived from 4 hospitals in China. This study was conducted in accordance with the Declaration of Helsinki and was approved by the institutional review board of Peking University Cancer Hospital; the informed consent was waived.

The medical records from the 4 hospitals were searched from May 2011 to June 2018. All resected NF‑PNETs with definite pathologically confirmed LN status were enrolled, and the patients were excluded according to the following exclusion criteria: (1) no MRI available or MR images were not sufficient to analysis; (2) the time interval between MRI and surgery was more than two weeks; (3) patient received local or systemic treatment before surgery. The recruitment pathway is shown in Fig. 1. A total of 66 patients from Beijing Cancer Hospital constituted the training group, whereas 121 patients from Peking University First Hospital (n = 46), the Affiliated Hospital of Qingdao University (n = 37) and First Affiliated Hospital of Kunming Medical University (n = 38) constituted the validation group.

Flowchart of the study of the enrolled patients

Baseline clinical information consisting of gender, age, body mass index (BMI), symptom (present or absent), total bilirubin (TB), alanine aminotransferase (ALT), aspartate aminotransferase (AST), fasting blood glucose (FBG), neutrophil–lymphocyte ratio (NLR = lymphocyte count/ neutrophil count), carcinoembryonic antigen (CEA), carbohydrate antigen 199 (CA199), carbohydrate antigen 724 (CA724), and neuron-specific enolase (NSE) were acquired from the medical records. Pathological analysis was based on the WHO 2019 classification, including the tumor grade (according to mitotic count and Ki-67 index), vascular invasion, neural invasion, and lymph node status.

All examinations were performed on 1.5 T (n = 37) or 3.0 T (n = 150) MRI scanners, using an 8-channel phased array body coil with the patients in the supine position. The MRI sequences included T2-weighted single-shot fat spin echo (SSFSE), FSE T1- weighted imaging and DWI. DWI was performed with single-shot echo-planar imaging (EPI) sequence prior to contrast administration with at least with b value of 0 and 1000 s/mm2. Dynamic Contrast-enhanced MRI was performed using a breath-hold fat-suppressed 3D T1-weighted LAVA-Flex sequence before and after intravenous administration Gd-DTPA (Magnevist, Bayer Schering Pharma, Berlin, Germany) at a dose of 0.1 mmol/kg and 2 mL/s, followed by a 20 mL of saline solution flush using a power injector. Images were acquired in arterial phase (20–35 s), portal phase (60–80 s), and delayed phase (180–240 s), respectively. The MR protocols are listed in Additional file 1: Table S1.

All the patients were distributed in random order, and the reviewers were blinded to the clinical information and the pathological reports.

Two radiologists evaluated the qualitative variables independently (H.B.Z. and P.N., both with 12 years' experience in abdominal MRI), and inter-observer agreement was evaluated. When there was a discrepancy, a senior radiologist (X.Y.Z., with 15 years' experience in abdominal MRI) was introduced for arbitration, and the result after arbitration was used in next analysis. The following qualitative features were evaluated: (a) tumor location (pancreatic head/neck, body or tail), (b) size (maximal axial dimension), (c) signal intensity (SI) on T2-WI (hypointense, isointense, or hyperintense relative to the surrounding pancreatic parenchyma), (d) exophytic growth (present or absent), (e) hyperenhancement at arterial phase (present or absent), (f) presence of upstream common bile duct dilatation (CBDD) and/or main pancreatic ductal dilatation (MPDD) due to tumor compression, (g) presence of vascular and adjacent organs invasion, (h) presence of synchronous liver metastases, (i) tumor margin (regular or irregular). The distal main pancreatic duct of the tumor was considered dilated when its diameter was ≥ 5 mm, while common bile duct dilatation was defined when its diameter was ≥ 10 mm. Vascular invasion was defined as the tumor directly invaded adjacent vessels with the results of lumen obstruction or occlusion, abutted more than 90° of major peripancreatic arteries, or abutted more than 180° of the adjacent vein. Regular margin was defined as: the round or oval shape with clear demarcation (Fig. 2a–f). Otherwise, the tumor with extra-nodular growth and confluent multinodular growth were defined as irregular margin (Fig. 3a–f) [23, 24]. The interobserver level of agreement for tumor margin was assessed by two blinded radiologists independently.

Pancreatic neuroendocrine tumors with regular tumor characteristics on T2WI (left), DWI (middle) and arterial phase (right) images. a–c a round, well-demarcated tumor with smooth contours is shown on MR images. The tumor-pancreas interface is sharp with pseudo-capsule (arrow). d–f MR images reveal oval but regular tumor located at the head of pancreas, the tumor-pancreas interface is clear and smooth. Although the pseudo-capsule is not demonstrated on MR images, the tumor is also categorized with regular shape

NF-PNETs reveal irregular characteristics on T2WI (left), DWI (middle) and arterial phase (right) images. The tumor shows ill-defined nodular tumor–pancreas interface with infiltrative to adjacent normal pancreatic parenchyma on (a–c). Highly infiltrative tumor with confluent multinodular growth and lacking demarcation is detected on (d–f), showing directly penetration to the duodenum with pathologic proven

If the tumor was located in the pancreatic head/neck, regional nodes included those along the common bile duct, common hepatic artery, portal vein, the anterior and posterior surfaces of the pancreatic head, and along the superior mesenteric artery. If the tumor was located in the pancreatic body/tail, regional nodes included those along the common hepatic artery, splenic and superior mesenteric artery [25]. All visible regional lymph nodes in the field of scan were analyzed. The size of the largest lymph node (the long axis and short axis) was measured, and short/long ratio was calculated subsequently. The number of the lymph nodes with the short axis > 5 mm, > 10 mm detected on DWI sequence was also recorded. In addition, morphological involvement of LNM was reported when the lymph node with abnormal round morphology or central necrosis.

Regions of interests (ROIs) were manually placed on the DWI images with b value of 1000 s/mm2 by two radiologists working together. DCE-MRI and T2WI images were used as reference for ROI segmentation. ROIs were also drawn long the outer border of primary pancreatic tumor on every slice with carefully avoiding vascular structures, biliary duct, pancreatic duct and normal pancreatic tissue. ADC values from whole slices of the lesion were averaged as the ADCmean. The maximum (ADCmax) and minimum (ADCmin) ADC value of the tumor were also recorded. Tumor volume was then multiplied by the slice thickness.

Routine examinations, including radiography and laboratory tests, were performed every 3–6 months for the first 2 years and then annually. Disease-free survival (DFS) was defined as the interval between operation and an event (tumor recurrence, death or last negative follow-up). The last date for follow-up was June 27, 2021.

The differences of the clinical factors and MRI features between LNM and non-LNM groups were compared by using independent t test or Mann–Whitney test for continuous variables and Chi-square test or Fisher’s exact test for categorical variables. Interobserver agreement was evaluated using Kappa coefficient, 0.0–0.20, 0.21–0.40, 0.41–0.60, 0.61–0.80 and 0.81–1.00 was considered slight, fair, moderate, substantial and perfect agreement. Univariate logistic regression included 33 variables according to LNM status; Bonferroni correction was used for multiple comparison (p < 0.05/33 ≈ 0.0015 was considered as statistically significant, p < 0.01 was considered as potentially significant for univariate logistic regression). Multivariate logistic regression model was established by substituting potentially significant variables from univariate analysis into equation. Independent factors associated with LNM were tested with odds ratios (OR) calculated and then, used for establishing a multivariate model for predicting LNM in NF-PNETs. The diagnostic performance of the model was evaluated by receiver operator characteristic (ROC) curve with area under ROC (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and accuracy calculated in both the training and validation cohorts. The cutoff value was selected using the maximum Youden’s index. Nomogram was yielded for clinical application. DFS curves of model-defined LNM groups and pathology LNM groups in the validation group were compared using Kaplan–Meier method with log-rank estimates. All statistical analyses were performed with IBM SPSS (Version 22.0; IBM Corp., New York, USA), and R package 3.6.2 (R Foundation for Statistical Computing, Vienna, Austria) Two-sided p-value < 0.05 was considered as statistically significant.