The study protocol was in complies with the Declaration of Helsinki and acts in accordance to ICH GCP guidelines. Institutional review board of Affiliated Hospital of Nanjing University of Chinese Medicine, Jiangsu Province Hospital of Chinese Medicine approved this study and explicitly waived the informed consent due to its retrospective nature. It also clarified that authors had access to identifying patient information when analyzing the data.
Patients
From January 2018 to December 2022, a total of 213 patients with a clinical diagnosis of pancreatic cancer were reviewed. The inclusion criteria were: (a) pathologically confirmed as PDAC and chemotherapy was performed; (b) patients were classified as unresectable disease according to the National Comprehensive Cancer Network (NCCN) guideline or present with metastatic disease; (c) preoperative contrast-enhanced CT performed within two weeks before therapy; The exclusion criteria were: (a) with a history of malignancy or combined with other malignancy (n = 7); (b) with pancreatic surgery history or with any local/systematic treatment before such as radiotherapy and chemotherapy (n =46); (c) the follow-up period was less than three months and no outcome events occurred (n=14 )
A total of 146 patients, including 80 men and 66 women, with a median age of 64 years (range, 42 – 88), were enrolled in the study. Then, 102 and 44 patients were allocated to the training and testing cohorts using random sampling at a fixed ratio of 7:3, respectively.
CT examination
All patients underwent contrast-enhanced CT scans on a multidetector CT scanner (brilliance iCT 256, Philips, The Netherlands). The scan ranged from the diaphragm to the pubic symphysis. After a plain CT scan, patients received 1mL/kg body weight of contrast media (Iopromide 370mg I/mL) at a rate of 3.5mL/s through the elbow. Dual-phase enhanced scanning was performed using mass injection tracking technology. An area of interest was set in the abdominal aorta. The CT value reached 100 HU and automatically triggered arterial phase scanning. Portal vein phase scanning was performed 60 to 70 seconds after triggering. The scanning parameters were as follows: tube voltage, 120 kVp; tube current, automatic milliampere technology; slice thickness, 2mm; layer spacing, 2mm; pitch, 1.0; FOV, 320mm × 320mm, collimator width, 128 mm × 0.6 mm; matrix size, 512 × 512.
The medium interval between CT examination and chemotherapy was 7 days (range, 2- 13).
Region-of-interest segmentation and radiomics features extraction
The loading and display of CT images, delineation of the volume of interest, and feature extraction were performed using 3D Slicer (version 4.11.1; http:// www. slicer.org), an open-source medical image analysis and visualization platform.
Two radiologists (with 8 and 10 years of experience in abdominal imaging diagnosis, respectively), who were blinded to the survival status of patients, delineated the boundaries of the tumor lesion independently. Region of interest (ROI) was defined as the area along the edge of each lesion layer of the tumor manually delineated on the portal vein phase transverse image. ROIs sometimes included visible necrosis and large blood vessels within the tumor, excluding adjacent pancreatic parenchyma.
The software automatically reads the CT value of each pixel within the volume of interest (VOI) (mean area of VOIs, 26.86cm³± 22.80; range 2.13–186.95 cm³), generating 855 parameters including 4 types: first-order statistics, shape-based feature, texture feature, wavelet-based feature. Image preprocessing and feature extraction were performed using the open-source Pyradiomics package (version 2.2.0: http:// www. radio mics. io/ pyradiomics. html). The voxel spacing was standardized with the size of 1 × 1 × 1 mm and voxel intensity values were discretized with a bin width of 25 HU to reduce the interference of image noise and normalize intensities.
Clinicopathologic data collection, chemotherapy, and follow-up
For all patients, clinical symptoms, preoperative biochemical data, and pathology related information were collected from the medical records. Clinical manifestations were documented as abdominal pain (without/with), jaundice (without/with), and diabetes (with/without). The main biochemical data included total bilirubin (TB), albumin (ALB), and CA19-9. The pathology related information, including the clinical-TNM (cTNM) stage according to the American Joint Committee on Cancer (AJCC) 8th edition staging system [17] and major peripancreatic vascular invasion status, was recorded according to preoperative CT imaging.
All patients underwent endoscopic ultrasound-guided tumor needle biopsy, and chemotherapy treatment was performed after pathological results were obtained. All patients received concurrent gemcitabine combined with a platinum agent, erlotinib, or a fluoropyrimidine. After treatment, patients were followed in the outpatient clinic every 3 months for the first years and every 6 months thereafter. During a median follow-up time of 8.0 months (mean, 10.3, range, 2.7 – 31.3), 38 patients were still alive, 99 patients died, and 9 patients were lost to follow-up.
Statistical analysis
Categorical variables were compared using χ2 validation or Wilcoxon validation. Continuous variables were compared by Student t validation or Mann-Whitney U validation, when appropriate. Interobserver agreement of radiomics features between two radiologists was assessed with intraclass correlation coefficients (≥ 0.75 were kept for further analysis). Radiomics features were selected by using a parametric method, the least absolute shrinkage and selection operator (LASSO) Cox regression [18,19]. Based on the radiomics features and clinicopathological data, univariate and multivariate Cox regression were performed, and we constructed the radiomics model, clinicopathologic model, and combined model. Feature selection and model building were carried out in the training cohort, whereas model performance was examined in the testing cohort. According to the median predicted risk score, the training cohort was divided into two groups (high-risk group versus low-risk group). We subsequently applied the same cutoff to the test cohort. The log-rank test was used to compare the overall survival between groups. The prognostic performance of each model was evaluated using AUC according to the predicted risk. Comparisons between the three models were performed using the Delong validation. A p-value < 0.05 (two-tailed) was considered statistically significant. Statistical analyses were performed with SPSS, version 22.0 (IBM, Armonk, NY, USA) and R software (R Foundation for Statistical Computing, version 3.4.1; https://www.r-project.org/).