This retrospective study was approved by the Biomedical Research Ethics Committee of our institution, and the need to obtain informed consent from patients was waived. Data were collected on consecutive patients with histopathologically diagnosed PASC and PDAC between January 2011 and December 2020. A total of 544 patients diagnosed with PASC and PDAC met the inclusion criteria, which were (1): histopathologically confirmed PDAC and PASC and (2) having undergone enhanced MRI examination within 30 days of surgery. The exclusion criteria were as follows: (1) a history of squamous cell carcinoma in other organs and (2) having undergone neoadjuvant chemotherapy before imaging examination. The inclusion and exclusion criteria are presented in Figure 1 . This study fundamentally followed the guidelines in the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) consensus (19). It should be noted that this study included the patients who all underwent enhanced MRI, because the performance of ring-enhancement was compared with the radiomics model.

All patients in this study underwent contrast-enhanced MRI of the pancreas using a 3.0-T system (Signa HDxt MR750 3.0-T, GE Healthcare, Milwaukee, USA; Skyra 3.0-T, Siemens, Erlangen, Germany). All patients were positioned supine, with a phased-array receiver coil covering the upper abdomen. The breath-hold single-shot fast-spin echo cross-sectional T2W sequence (repetition time/echo time [TR/TE], 6316/87 ms; field of view [FOV], 360 × 420 mm²; matrix, 224 × 270; slice thickness, 5 mm; slice gap, 1 mm) and unenhanced cross-sectional T1-weighted(T1W) fat-suppressed sequence (TR/TE = 2.58/1.18 ms; FOV = 440 × 440 mm²; matrix = 224 × 270; slice thickness = 5 mm; no slice gap) were used. Dynamic contrast-enhanced images, including the arterial phase (15 s), pancreatic parenchymal phase (20 s), and portal venous phase (40 s) images, were obtained using a fat-suppressed sequence (volumetric interpolated breath-hold examination; TR/TE=3.0/1.3 ms; FOV = 440 × 440 mm²; slice thickness = 5 mm; matrix = 224 × 270) after intravenous injection of gadopentetate dimeglumine (Magnevist and Gadovist, Bayer Schering Pharma, Berlin, Germany). A contrast agent dose of 0.2 mL/kg was administered intravenously at a flow rate of 2 mL/s, followed by 20 mL of normal saline (to flush the tube).

Cross-sectional T1WI, T2WI and dynamic contrast-enhanced images were used for analysis. All images were analyzed by two abdominal radiologists (YB and QL, with 20 and 10 years of experience, respectively) blinded to the clinical and pathological details. Notably, the discordance between the two radiologists was resolved by consensus.

On the basis of a comprehensive literature research and our local experience on PASC and PDAC, the following characteristics were assessed (1): MRI-reported tumor size (the maximum cross-sectional diameter of the tumor (20)); (2) tumor location (head/uncinate or body/tail (21)); (3) pancreatitis (peripancreatic fluid (22)); (4) diameter of main pancreatic duct (MPD); (5) common bile duct (CBD) cutoff and dilation (> 10 mm (23)); (6) cyst (presence of pseudocysts and retention cysts with high signal intensity on T2-weighted (T2W) image (24)); (7) contour abnormality (bulging contours or loss of normal lobulation (25)); (8) parenchymal atrophy (anteroposterior body diameter less than 20 mm (23)); (9) MRI-reported lymph node (LN) metastasis (LN short-axis diameter > 10 mm (21)); and (10) ring-enhancement (26).

The MRIs of all patients were loaded into three-dimensional (3D) Slicer version 4.8.1 (an open-source software; https://www.slicer.org/), and regions of interest were manually drawn slice-by-slice on T1WI and T2WI (QL). A senior radiologist (YB) delineated all patients. We trained an automatic tumor segmentation model in the training set. Then this automatic segmentation model was applied in the validation set. The mask of the automatic segmentations was compared with the mask delineated by the senior radiologist using dice similarity coefficients (DSCs) in the validation set.

Moreover, 50 randomly selected patients including 25 patients with PDAC and 25 patients with PASC were used for delineation studies of pancreatic tumors (QL and WBL) for interobserver and intraobserver DSCs. Specifically, delineation was conducted two weeks apart in the intraobserver study.

The segmentation network of pancreatic tumors using nnU-Net, a robust adaptive framework based on 2D U-Net and 3D U-Net, was used for pixel-level semantic segmentation of NII data. In addition, nnU-Net automatically generated network structures and hyperparameters according to the characteristics of the data. The input patch size was 64 × 192 × 192, with a batch size of two. A total of five downsampling operations was performed, which resulted in a feature map size of 2 × 6 × 6 in the bottleneck. The initial number of convolutional kernels was set to 32, which was doubled with each down-sampling to a maximum of 320. The number of kernels in the decoder mirrors was in the encoder. Leaky rectified linear units (leaky ReLUs) were used as the nonlinear activation functions. Instance normalization was used for the feature map normalization. The training objective was the sum of the dice and cross-entropy losses.

Where P is the predicted segmentation result, and T is the labeled segmentation result.

y_i indicates the label value, and y i  ′ indicates the predictive value

Loss operated in one class, which labels the pancreatic tumor. The nnU-Net used stochastic gradient descent with an initial learning rate of 0.01 and a Nesterov momentum of 0.99. Training runs were conducted for 1000 epochs, where one epoch was defined as 250 iterations. The learning rate decayed using a polynomial schedule. Subsequently, training patches were cropped from randomly selected training cases.

The network was fully trained using labeled MR images (382 for T1WI and T2WI). We tested the performance of the segmented network on the validation set (the same patients as the above-mentioned validation set). The nnU-Net is a free and open-source out-of-the-box segmentation tool, and the source code is publicly available on GitHub (https://github.com/MRCWirtz/nnUNet-1, accessed on April 12, 2022). The software requires only a set of annotated MRIs as input data and a mainstream computer with a powerful GPU.

Normal distribution and variance homogeneity tests were performed for all the continuous variables. The Student’s t-test (normal distribution), Kruskal–Wallis H test (skewed distribution), and chi-square test (categorical variables) were used to compare the differences between the two groups. The interobserver DSCs between the automatic segmentation and manual delineation of tumors were reported. Interobserver agreement was quantified using the k statistic for categorical variables, which was graded as 0–0.20, poor agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.80, good agreement; and 0.81–1.00, excellent agreement, and the intraclass correlation coefficient (ICC) for continuous variables, which was graded as 0–0.49, poor agreement; 0.50–0.75, moderate agreement; 0.76–0.90, good agreement; 0.91–1.00, excellent agreement. Deaths were defined as events, and deaths attributed to other causes were set as censored observations. Overall survival (OS) was calculated from the surgery date to the time of death or the end of the follow-up period.

Furthermore, univariate logistic regression analysis was used to select the significant clinical and radiomic characteristics (p< 0.05). A multivariable logistic regression analysis was conducted to develop a model for the differential diagnosis of PASC and PDAC, and a nomogram was constructed. We constructed a multivariable model using a stepwise regression method based on the Akaike information criterion to determine the best-fitting model (30). Subsequently, the discrimination of the model was evaluated using a receiver operating characteristic (ROC) curve, and the area under the ROC curve (AUC) was calculated concurrently. ROC curves were compared using the DeLong test. Additionally, the sensitivity, specificity and accuracy of the models were reported.

In this study, we developed two prediction models as follows: MRI and radiomics models. The MRI model was developed using conventional MRI characteristics, whereas the radiomics model was developed using conventional MRI characteristics and rad-score. The clinical usefulness of the models was tested using decision curve analysis (DCA) by quantifying the net benefit at different threshold probabilities. Finally, Kaplan–Meier estimates were applied to graph the survival curves and the log-rank test was used to analyze the differences between the predicted PASC group and the predicted PDAC group.

Statistical significance was defined as a two-tailed p-value< 0.05. All analyses were performed using the R software (version 3.3.3, R Foundation for Statistical Computing).

A tatal of 510 patinets were included in this study. Of the 510 included patients, 387 consecutive patients (age: 61 ± 9 years; range: 28–86 years; 250 males) had PDAC and 123 consecutive patients (age: 62 ± 10 years; range: 36–84 years; 78 males) had PASC. All patients were divided into training and validation sets according TRIPOD guidelines (19). The prediction model was developed on the training set of 382 consecutive patients (age: 61 ± 9.9 years; range: 28–85 years; 239 males), enrolled between January 2011 and February 2018. Thus, the validation set consisted of 128 consecutive patients (age: 60 ± 10 years; range: 35–83 years; 89 males), enrolled between March 2018 and December 2020. Interobserver agreements between two radiologists for MRI characteristics were excellent, and k statistic ranged from 0.91 to 0.93. The interobserver ICCs of MRI-reported tumor size were also good, ranging from 0.82 to 0.88. Significant differences were observed in tumor location, MRI-reported tumor size, CBD cutoff and dilation, diameter of MPD, and ring-enhancement in the training and validation sets. The BMI and MRI reported that LN status was significantly different in the training set. Table 1 presents the summary of patient characteristics.

The intraobserver and interobserver DSCs for manual segmentation were both moderate, ranging from 0.73 to 0.80 and 0.71 to 0.76, respectively, in the T1WI and T2WI sequences. Additionally, the DSCs between the manual segmentation and automatic segmentation were moderate, ranging from 0.77 to 0.81 in the two sequences for the validation set.

We excluded radiomics features with non-significant differences between the groups or non-significant correlations between the two types of PC. Subsequently, the radiomics features were reduced to 328. Finally, the radiomics features were reduced to five by LASSO regression, and the LASSO formula was used to obtain the rad-score ( Table 2 ). Figure 3 shows the LASSO results for optimal features.

The AUC, sensitivity, specificity, and accuracy of the ring-enhancement were 0.80 (95% confidence interval [CI]: 0.74–0.85), 75.51%, 83.80%, and 81.68%, respectively, in the training set and 0.74 (95% CI: 0.62–0.86), 60.00%, 87.38%, and 82.03%, respectively in the validation set.

Multivariable logistic regression analysis included MRI-reported tumor size, tumor location, diameter of MPD, and CBD cutoff and dilation. Consequently, MRI-reported tumor size (p< 0.001), diameter of MPD (p = 0.037), and CBD cutoff and dilation (p = 0.014) were selected for the MRI model ( Table 3 ). In the MRI model, the AUC, sensitivity, specificity, and accuracy were 0.71 (95% CI: 0.65–0.77), 67.35%, 69.01%, and 68.59%, respectively, in the training set and 0.80 (95% CI: 0.70–0.89), 84.00%, 64.08%, and 67.97%, respectively, in the validation set.

A total of 195 patients were accurately identified among 284 patients (69%, 195 of 284) in the PDAC group, whereas 88 patients (90%, 88 of 98) were accurately identified among 98 patients in the PASC group using the radiomics model in the training set ( Figure 4A ). In contrast, 87 patients were accurately predicted among 103 patients (84%, 87 of 103) in the PDAC group, and 20 patients (80%, 20 of 25) were accurately identified among 25 patients in the PASC group using the radiomics model in the validation set. ( Figure 4B ). The median follow-up duration was 18.7 months (IQR, 9.4–34.4 months) in the PDAC group and 13.7 months (IQR, 7.2–21.9 months) in the PASC group. During the observation period, 283 and 85 patients in the PDAC and 85 in the PASC groups died, respectively. A log-rank test revealed a longer survival duration in the predicted PDAC group than that in the predicted PASC group (training set: p = 0.018, validation set: p = 0.001) ( Figure 4C, D ).

Furthermore, the multivariable logistic regression analysis included MRI-reported tumor size, tumor location, MPD diameter, CBD cutoff and dilation, and rad-score. Consequently, MRI-reported tumor size (p = 0.007), diameter of MPD (p = 0.035), CBD cutoff and dilation (p = 0.057), and rad-score (p< 0.0001) were selected for the radiomics model. In the radiomics model, the AUC, sensitivity, specificity, and accuracy were 0.86 (95% CI: 0.82–0.90), 89.80%, 68.66%, and 74.08% in the training set and 0.87 (95% CI: 0.79–0.96), 80.00%, 84.47%, and 83.59% in the validation set ( Figure 4A, B ). There was a significant difference in AUCs between the MRI model and radiomics model (p< 0.001) according to the DeLong test in the training and validation sets. The results of the two multivariable logistic regression models are shown in Table 3 . Figure 5 shows the two cases using the two nomograms. The performances of all prediction models are shown in Table 4 and Figure 6A, B .

The DCA showed that if the threshold probability were 0.05–0.81 and 0.02–0.78 in the training and the validation sets, respectively, using the radiomics model to distinguish PASC from PDAC was more beneficial than the treat-all-patients as a PASC scheme or the treat-all-patients as a PDAC scheme ( Figure 6C, D ).

Our study had some limitations. First, this was a retrospective and single-center study with potential for bias. Second, an independent external validation dataset for another center will be needed in future study. Third, although our study only included sequences from non-enhanced MRI, the diffusion-weighted imaging and apparent diffusion coefficient should be investigated in the future to improve our model. Lastly, the postoperative information in many cases was unavailable, leaving potential room for bias, because patients were often followed up by the local oncologists.