This was a single-institution, retrospective, diagnostic, descriptive study. Cumulative data were collected from August 2016 to October 2020 at Sun Yat-sen University Cancer Center, Guangzhou, China, for 953 patients with PDAC who received CT scan or MR scan. The regional lymph nodes were matched with the pathological results, and the patients were divided into two groups: a metastatic lymph node group (n = 99) and a benign lymph node group (n = 100). Based on the preoperative venous phase CT images, the critical inclusion criteria were: 1) pathologically confirmed PDAC and LNM and 2) DECT scanning within 1 month before surgical resection. Exclusion criteria were: 1) history of any systemic therapy before surgical resection and 2) missing preoperative clinical and image data. The complete patient enrollment process is shown in Fig. 1. This study was approved by the institutional ethics committee of Sun Yat-sen University Cancer Center, Guangzhou, China (IRB number: B2019-012-01). Informed consent was waived based on the retrospective study design.

Flowchart of all patients

All patients were scanned using a dual-source CT scanner (SOMATOM Force, Siemens Healthcare) with the dual-energy mode. The acquisition parameters were detector collimation set at 192 × 0.6 mm, and 2) tube voltage and tube current set as follows: pitch 0.9, A/B tube current 250/125 mAs, and tube voltage 100 kV/Sn 150 kV. The CARE Dose 4D automatic exposure control system was turned on. The contrast agent iopromide 370 (370 mg/mL iodine, Bayer, Mississauga, Ontario, Canada) was intravenously injected through the antecubital vein by a power injector (XD 8003 Ulrich) with a 20-gauge needle. Contrast agent application was controlled using the bolus-tracking technique in the descending aorta (signal attenuation threshold, 100 HU). Data acquisition was initiated after the threshold was reached in the abdominal aorta, with a mean delay of 7 seconds. After 20 mL of saline solution was injected at flow rates of 4.0 mL/second, 1.5 mL/kg of contrast agent was used, followed by 40 mL of saline solution at flow rates of 4.0 mL/second. After initiation of contrast agent injection, multiphasic scanning was started with a 30- to 35-second delay for the arterial phase, a 60- to 65-second delay for the portal phase, and a 180-second delay for the delayed phase.

All the CT images were transferred to an off-line workstation (Siemens syngo.via VB20 software). Venous phase images were selected to reconstruct virtual monoenergetic images. We developed and standardized ROI outlining criteria in our study. Manual outlining of the major part (about 60%) of the solid portion of the lymph node was performed. And we prevented drawing two small ROI to avoid potential low signal-noise rate. We also keep the ROI always in the solid portion of the LN to avoid partial volume effect. All measurements were performed twice on two consecutive maximal slices in the selected lymph nodes, and their average values were calculated. The slope of the spectral Hounsfield unit curve (in Hounsfield unit per kiloelectron-volt [HU per KV]) on virtual monochromatic imaging, defined as the difference between the CT value at 40 keV and that at 70 keV divided by the energy difference (30 keV), was calculated as follows:

$$\lambda HU=\left(HU \,40 \,keV-HU \,70 \,keV\right) / 30$$

The HU 40 keV and HU 70 keV values represent the CT values measured on 40 keV and 70 keV images, respectively. Dual-energy parameters were measured by placing a region of interest such that it encompassed the solid portion of lymph nodes. The electron density (Rho), effective atomic number (Z), iodine concentration (IC), normalized iodine concentration (NIC), and the dual energy index (DEI) were obtained, and the differences were compared between the parameters for benign and metastatic lymph nodes. All measurements were performed twice on two consecutive maximal slices in the selected lymph nodes, and their average values were calculated. For cases in which the location was unclear, a radiologist-defined region of interest was determined based on pathological or surgical records (Fig. 2). The radiologists did not know the pathologic findings of the lymph nodes when they measured the energy spectrum parameters for diagnosis.

Measurement of dual-energy computed tomography (DECT) parameters in the same patients for different lymph nodes. The left column shows benign lymph nodes; the right column shows metastatic lymph nodes. Venous phase contrast-enhanced DECT images show the target lymph node with the following characteristics: iodine-based pseudo-colorized images for A a benign lymph node and B metastatic lymph node; representative effective atomic number images for C a benign lymph node and D a metastatic lymph node; representative venous phase CT images for E a benign lymph node and F a malignant lymph node; and graph of the spectral HU curves for G a benign lymph node, and H a malignant lymph node. CT, computed tomography; HU, Hounsfield unit

LN specimens were harvested by two pancreatic surgeon and the whole LN was fixed for HE staining with 5-um thin serial section. Surgical team members marked the resected specimen to dissect the lymph nodes. After the surgery, a microscope was used to show the regional LNM. Pathological results of included lymph nodes were used for gold standard by two pathologists with more than 6 years of experience. Pathology reports were obtained and reviewed for the patient sex and age and the tumor characteristics of location, size, differentiation status, growth pattern, and tumor–node–metastasis stage.

All regional lymph nodes in pancreatic ductal adenocarcinoma were divided into many levels in our hospital to ensure that the LNs selected on CT images corresponded to the pathological results, and surgeons usually strived to resect lymph node stations numbers 5, 6, 8a, 12, 13a, 13b, 14a, 14b, and so on (AJCC 8th edition). We collaborated with pathologists and surgeons. Before LNs dissection, target LNs were selected for spectral parameter measurements. We have adopted the following methods to ensure that the lymph nodes selected on abdominal CT images corresponded to the pathological results. We usually select lymph nodes in a specific station or region on preoperative CT. The size, shape, and three-dimensional location of the target LN in one node station were marked and recorded preoperatively on CT images by radiologists. We used coronal, axial, and sagittal CT images to help allocate LN. We measure its size on CT images and mark its relationship with blood vessels and nearby organ. During the surgery, the surgeon removes lymph node region by region and then labeled them. These targeted lymph nodes were put into some small clear bags with formalin solution separately, and the pathologist performed the diagnosis of the marked lymph node according to the lymph node station and records. It was shown that how we identified LNs in histopathology, as Supplemental Fig. 1. For lymph node in one known station, the following two methods have been further used in our study. First, according to the location (such as the subgroup, or nearby vessels), internal characteristics, shape and size of lymph nodes, doctors cooperate to match and correlate the size and morphological characteristics of the target lymph nodes under CT and microscope imaging. Second, if all lymph nodes in one region or station are malignant or benign, one node is selected for measurement- usually the largest one, for imaging-pathological correspondence.

Statistical analyses were carried out using the statistical software SPSS, version 26.0 (SPSS) and MedCalc, version 19.3 (MedCalc Software). Continuous variables were presented as mean ± standard deviation (SD). Outliers were detected using box plots. The mean of the duplicate measurements was used in statistical analysis for all samples. The independent sample t-test or Mann–Whitney U test was used to compare continuous variables. The level of significance was set at p ≤ 0.05. The ROC curve, the areas under the ROC curve (AUC), 95% confidence interval (CI), sensitivity, specificity, Youden index, cutoff point, positive predictive value, and negative predictive value of single DECT parameter models and multiparameter diagnostic models were obtained using MedCalc, version 19.3, software, and these parameter models and multiparameter diagnostic models were also used to assess performance. The AUCs were compared using the Z test. A value of p < 0.05 was considered statistically significant.

The DeLong test was used to compare AUCs. A p value of <0.05 was considered statistically significant. Binary logistic regression analysis using a forward stepwise (conditional) method was performed. A regression model was established by combining multiple indicators with statistically significant (p values < 0.05), and variables with p values > 0.1 in univariate logistic regression models were excluded from the multivariable logistic regression models. This approach was based on research by Hosmer and Lemeshow (1989) on a goodness-of-fit test for the model, with a p value > 0.05 considered a good fit.

All diagnostic indicators were used to analyze the difference of the diagnostic value between benign and metastatic lymph nodes. The diagnostic efficacy of IC, predictive probability, and other indicators for lymph nodes < 5 mm and > 5 mm were compared, and the diagnostic efficacy of these indicators was compared for benign and metastatic lymph nodes.