Patient population

This study received approval from the local institutional ethics committee, and the requirement for written informed consent was waived for the retrospective nature. AP patients were retrospectively screened from August 2015 to March 2022. The flowchart of patient recruitment is shown in Fig. 1. The inclusion criteria for the study were as follows: (1) patients were diagnosed with AP according to the revised Atlanta classification and definitions (2012 version) [14], who underwent upper abdominal contrast-enhanced CT scan within 72 h of symptom onset; (2) all patients were treated with fasting, acid and enzyme suppression, fluid infusion and laboratory test obtained during hospitalization; and (3) patients underwent subsequent abdominal follow-up contrast-enhanced CT scan at 7–10 days after initial CT scan according to the revised Atlanta classification and definitions (2012 version) [14]. The exclusion criteria consisted of: (1) patients were diagnosed with concomitant tumors or chronic wasting disease; (2) patients with a history of chronic pancreatitis were excluded due to the potential assessment bias induced by chronic interstitial changes (excluding acute onset of chronic pancreatitis); (3) patients with incomplete clinical data; (4) CT images of poor quality; and (5) patients with age < 18 years. The various clinical characteristics of patients, gender, age, etiology, history of diabetes, severity of AP, Bedside Index for Severity in Acute Pancreatitis (BISAP) score, white blood cell (WBC), C-reactive protein (CRP), procalcitonin (PCT) were collected. Further, the BISAP score calculation and laboratory test were conducted on all patients within 24 h of admission. The severity of AP was categorized into three degrees: mild (MAP), moderately severe (MSAP), and severe AP (SAP), in accordance with the revised Atlanta classification [14].

Fig. 1

Flowchart of patient enrollment in the current study

Definition of prognosis

Patients were categorized into good and poor prognosis subgroups mainly based on the follow-up CT scan indicating the occurrence of pancreatic and/or peripancreatic infected necrosis or persistent (≥ 48 h) organ failure according to the previous literature and our local experience [14,15,16,17]. Categorization was performed by two radiologists (H.C. and Y.W., with 4 and 3 years of experience, respectively). The presence of infection might be indicated when extraluminal gas is observed in pancreatic and/or peripancreatic tissues on contrast-enhanced CT [14]. Organ failure was defined using the modified Marshall scoring system, and a score of 2 or more in any system indicates the presence of organ failure [16, 18]. The poor prognosis group met one of the following conditions: (1) follow-up CT showed an increase in lesion size or modified CT severity index (MCTSI); (2) there is extraluminal gas in the pancreatic and/or peripancreatic tissues on follow-up CT indicating the presence of infected necrosis [14]; or (3) the occurrence of infection or persistent organ failure during hospitalization.

CT image acquisition

CT examinations were performed by using the commercial multidetector scanner (256-section Philips Brilliance iCT, Philips Medical Systems) covering from the diaphragm level to the inferior pole of the kidneys. The acquisition parameters for the CT images were as follows: tube voltage, 120 KV; tube current, auto mAs; layer thickness, 5 mm; layer spacing, 5 mm; pitch, 0.984–1.375; field of view, 300 × 400 mm; matrix size, 512 × 512. In addition, iohexol (350 mgI/mL) was administrated intravenously through a peripheral vein at a flow rate of 3.0–4.0 mL/s with a dose of 1.5 mL/kg [19], using a pressure syringe. The arterial phase images were obtained with a post-injection delay of 25–28 s, while the venous phase images were obtained with a post-injection delay of 60–70 s.

Image segmentation and radiomics extraction

Two experienced radiologists (X.L.W. and D.L., with 8 and 10 years of experience in the field of abdominal imaging, respectively) conducted a joint review of all abdominal CT images and reached a consensus after discussing if there was disagreement regarding the CT features. CT features were evaluated, including pancreatic enlargement, peripancreatic inflammation, peripancreatic effusion, peripancreatic gas, pancreatic necrosis, extrapancreatic complications such as pleural effusion, peritoneal effusion, vascular or gastrointestinal complications. All extrapancreatic complications were lumped together and considered as a single variable that was either present or absent. The CT images of the venous phase were imported into the uAI research portal (uRP) (version 211230), and radiomic features were extracted from regions of interest (ROIs) via this software. The pancreatic ROIs were manually delineated by the two radiologists on each axial slice along the edge of the pancreatic parenchyma covering the whole pancreatic region. The corresponding peripancreatic ROIs were delineated by expanding the pancreatic ROI by 5 mm towards the peripancreatic area (the peripancreatic area encompassed a 5 mm distance from the pancreatic surface, excluding the pancreatic parenchyma area, blood vessels, bile ducts, peripancreatic lymph nodes and organs) (Fig. 2). After the image segmentation, to mitigate potential influences stemming from image and radiomics features, all images were resampled using a linear interpolation algorithm to achieve voxel dimensions of 1 mm × 1 mm × 1 mm [20]. The time spent intermittently on the entire process of segmentation and resampling by the researchers amounted to approximately 2 months. Furthermore, the radiomics features were normalized by Z-score standardization.

Fig. 2

Example of regions of interest (ROIs) for acute pancreatitis. The pancreatic ROIs (blue area) were manually delineated along the edge of the pancreatic parenchyma, and the peripancreatic ROIs (red area) were delineated by expanding the pancreatic ROIs by 5 mm towards the peripancreatic area

Intra- and interobserver agreement

A set of 50 patient images was randomly selected to be assessed. Two radiologists (X.L.W. and D.L.) independently delineated pancreatic and peripancreatic ROIs to evaluate the interobserver agreement. To assess intraobserver agreement, a radiologist (X.L.W.) delineated these ROIs again with a gap of two weeks following the same procedure. The features obtained from the two extractions were then compared.

Radiomics and clinical feature selection and model construction

The selection of optimal radiomics features was performed through a sequential process involving three algorithms. Firstly, the variance threshold method was applied with a threshold set at 0.8, and then the Select K-Best method with ANOVA F-value (p < 0.05) was used to choose features. Finally, the least absolute shrinkage and selection operator (LASSO) method was utilized with parameters set at k-fold = 5 and alpha = 0.00236. Radscore was calculated by linearly multiplying the optimal radiomics features with their corresponding LASSO coefficients. The logistic regression classifiers were used to construct the models. The variance inflation factor (VIF) was used to assess the presence of multicollinearity among clinical variables and CT features. Variables with VIF > 10, including gender, severity of AP, pancreatic enlargement, and peripancreatic inflammation, were removed. The remaining variables were then included in the univariate and multivariate logistic regression analyses to identify independent risk factors associated with poor prognosis. Based on the independent risk factors and the optimal radiomics features, the clinical, pancreatic, peripancreatic, radiomics, and combined models were constructed. Additionally, the data of the test set was used to independently evaluate the performance of the combined model.

Statistical analysis

Statistical analyses were performed using SPSS (version 25.0; IBM) and R software (version 3.6.1). Normality testing of continuous variables was performed using the Shapiro–Wilk test. Continuous variables were compared by using the Student’s t-test or Mann–Whitney U-test. Categorical variables are presented as numbers and percentages. Categorical variables were compared using the chi-square test or Fisher exact test as appropriate. The area under the receiver operating characteristic curve (AUC) was calculated to evaluate the predictive ability. The calibration curve and the decision curve analyses (DCA) were used to evaluate the calibration and clinical practicability of the combined model using the R software. The AUCs of the five models were compared using the DeLong test. The inter-class correlation coefficient (ICC) was calculated by a two-way mixed-effects model to quantify the consistency of feature extraction [21], ICC > 0.75 in both test-retest and inter-reader analyses is considered indicative of good consistency. p < 0.05 was considered statistically significant.