Resectable pancreatic ductal adenocarcinoma: association between preoperative CT texture features and metastatic nodal involvement

The institutional review board approved this retrospective study and waived the written informed consent requirement. According to the database of pathology in 2016, a total of 245 consecutive patients with PDAC who were operated at our institution were enrolled. Patients underwent potentially curative resection (R0 and R1) with standard regional lymph node dissection. All PDAC were comfirmed by pathlogical analysis. The tumors were staged according to tumor-node-metastasis (TNM) cancer staging system (the 7th edition) by the American Joint Committee on Cancer (AJCC) and the International Union for Cancer Control (IUCC) [14]. The inclusion criteria were as follows: a. both primary tumor and regional lymph nodes surgical resections were performed and PDAC was confirmed by pathology; and b. contrast-enhanced CT images obtained on CT750 HD scanner (GE Healthcare, Milwaukee Wis, USA) using spectral imaging (GSI) mode; c. no neoadjuvant therapy before surgery. Finally, a total of 155 patients (104 men, 51 women; mean age 63.8 ± 9.6 years; range 37–83 years) were enrolled in this retrospective study (Fig. 1). Detailed pathological reports, including all resected specimens, were obtained through hospital information system (HIS). The total number of examinated lymph nodes and the number and location of the positive lymph nodes were recorded. Study subjects were divided into two groups according to lymph node status. Seventy-two cases with lymph node metastases (N+ group) and 83 cases without lymph node metastases (N- group) were used for texture analysis. According to the detail report of pathology, LNR of each case was also calculated in this study by the radiologist (XZL). LNR was calculated as the ratio of (positive number)/(total number). Among the 72 cases with positive lymph nodes, in 43 cases metastases were located in the peri-pancreatic region, in 16 cases peri-tumoral, and in 10 cases in the hepatoduodenal ligament (no.12), in 12 cases in no.14, in 6 cases in no.9, in 4 cases in no.16, in two cases in no.7 and no.5, and in one case in no.10 and no.6.

CT examinations were performed with a 64 row detector scanner (CT750 HD) using spectral imaging (GSI) mode with fast tube voltage switching between 80kVp and 140 kVp on adjacent views during a single rotation on the late arterial phase (AP) and portal venous phase (PVP). Contrast enhanced images were obtained after administration of nonionic contrast medium (Iopamidol Injection, Iopamiro 300; Shanghai BRACCO Sine Pharmaceutical Corp. Ltd., China. Ioversol, Optiray 320; Tyco Healthcare, Montreal, Quebec, Canada) at a dose of 450 mg iodine per kilogram of body weight. All the contrast medium was injected in 30 s, followed by 20 ml saline at the same injection rate. AP images were acquired at 35 s and PVP images were acquired at 65 s after the start of contrast material administration. CT images without contrast were obtained with a tube voltage of 120 kVp, an automatic tube current of 100–500 mA, a noise index (NI) of 10, and a rotation time of 0.6 s. CT images in AP and PVP were obtained with GSI mode (tube voltage switching instantaneously between 80 kVp and 140 kVp, and a dedicated GSI protocol with similar CT dose index (CTDI) of pre-contrast scan was selected. The other CT image acquisition parameters included: (1) pitch, 1.375; (2) detector collimation of 64*0.625 mm; (3) image matrix of 512× 512; (4) FOV (field of view) of 35~40 cm; (5) Virtual monochromatic (VMC) images were reconstructed with standard algorithm and slice thickness of 5 mm (at 65 k electron voltage, keV) and 1.25 mm (at 60 keV). An adaptive statistical iterative reconstruction (ASiR, 30% for 5 mm and 40% for 1.25 mm) algorithm was employed to reduce the image noise. The aforementioned protocol was the routine protocol for preoperative staging of pancreatic cancer in our hospital. Theoretically and according to the literature, virtual monochromatic image with lower keV level had higher contrast-to-noise-ratio (CNR) and noise [15, 16]. While only limited application of noise reduction software of ASiR in spectral imaging on CT750 HD scanner (GE Healthcare, Milwaukee Wis, USA) was approved by China Food and Drug Administration in 2016, the noise reduction technique could not be applied on the virtual monochromatic images lower than 60 keV in our CT scanner then. Considering the balance of CNR and noise, 60 keV and 65 keV were adopted for pancreas then in our center. 60 keV was also used in other studies [17, 18]. So there were totally 4 image subsets (5 mm-AP-65 keV, 5 mm-PVP-65 keV, 1.25 mm-AP-60 keV, 1.25 mm-PVP-60 keV) in this study, and each image subset used the same window level and window wide.

We conducted a texture analysis on the axial contrast enhanced VMC CT images on AP and PVP. Texture analysis was performed with MaZda statistical texture analysis software (version 4.6, available at http://​www.​eletel.​p.​lodz.​pl/​mazda/​). Images were reviewed through hospital PACS for all slices and the slice with maximum lesion diameter was exported as .bmp files. The ROI of tumor was drawn by two readers (WHF and XZL) with consensus, which takes about 15 min for one case. During this process, firstly, we choosed one of the four images with the lesion boundary most clearly displayed and drew the ROI manually on this image. Secondly, we copied and pasted the ROI on the other three images and adjusted accordingly. The two-dimensional (2D) region of interest (ROI) was manually placed within the boundary of tumor on each selected image (Fig. 2). In the difficult cases, in which tumors were obscure on CT, MRI was also referenced if it was available. Most patients suspicious of pancreatic tumor underwent both contrast enhanced CT and MRI before surgery in our hospital. There were 124/155 patients, who had preoperative MRI in this study. In addition, the maximum tumor diameter was also recorded for each case.

MaZda software allows computation of almost 300 texture features (Additional file 1) based on image histogram, co-occurrence matrix (COM), run-length matrix (RLM), image gradient, auto-regressive model (ARM) and Haar wavelet (WAV). All of these texture features were extracted from each ROI. To diminish the influence of contrast and brightness variation, we normalized the image grey level intensity in the range of μ-3δ, μ + 3δ (μ, mean grey level value; δ, mean standard deviation). A combination of feature selection algorithms (FPM) including Fisher’s coefficient (Fisher), classification error probability combined with the average correlation coefficients (POE + ACC), and mutual information (MI) was used to determine 30 texture features with the highest discriminative power for differentiation between lymph node metastases positive and negative groups.

The gender of the patients was compared between the two groups using Chi-square test. For the quantitative data with a normal distribution, which presented as mean ± standard deviation, independent samples t-test was used, and for the quantitative data with a skewed distribution, which presented as median (quartile 1, quartile 3), Mann-Whitney U test was used, to evaluate the difference between the two groups. Receiver operating characteristic (ROC) curves were generated for the features that presented significant differences, and the area under the ROC curve (AUC) was calculated and compared between different texture features. The 95% confidence intervals (CI) for AUC and p values for comparison of related ROC curves were obtained by the method described by DeLong and coworkers [19]. The relationship between the features (with AUC > 0.5) and LNR were also evaluated with non-parametric Spearman’s correlation coefficients. Values p < 0.05 were considered statistically significant. Statistical analysis was performed using SPSS software (SPSS version 23.0; SPSS Inc., Chicago, IL, USA) and MedCalc for Windows, version 11.4 (MedCalc Software, Ostend, Belgium).