Abstract
Purpose
To evaluate the value of a combined model based on multiphase contrast-enhanced CT radiomics and clinical features for differentiating hypovascular pancreatic neuroendocrine tumors (hypo-PNETs) from pancreatic ductal adenocarcinoma (PDAC).
Methods
A total of 297 patients with pathologically confirmed pancreatic tumors, including 99 hypo-PNETs and 198 PDACs, were retrospectively enrolled. Radiomics features were extracted from non-contrast, arterial-phase, venous-phase, and combined three-phase CT images. After feature selection, radiomics models were established using multiple machine-learning classifiers. Independent clinical predictors were identified by logistic regression and integrated with the optimal radiomics signature to construct a combined model. Model performance was assessed using receiver operating characteristic analysis, calibration curves, decision curve analysis, net reclassification improvement, and integrated discrimination improvement.
Results
Pancreatic duct dilatation, tumor composition, age, and maximum tumor diameter were identified as independent predictors. Among the radiomics models, the combined three-phase SVM model achieved the best performance, with AUCs of 0.814 and 0.812 in the training and test sets, respectively. The combined model yielded the highest AUCs (0.886 in the training set and 0.849 in the test set); however, because several between-model comparisons in the test set did not reach statistical significance, its advantage should be interpreted as a potential incremental benefit rather than definitive superiority.
Conclusion
A combined model integrating multiphasic CT radiomics and clinical features showed promising performance for differentiating hypo-PNETs from PDAC. This model may provide complementary support for preoperative diagnosis, although its incremental value over the radiomics-only model requires further validation.
Supplementary Information
The online version contains supplementary material available at 10.1007/s00432-026-06481-1.
Keywords: Pancreatic neuroendocrine tumors, Pancreatic ductal adenocarcinoma, Computed tomography, Radiomics, Machine learning
Introduction
Pancreatic ductal adenocarcinoma (PDAC) is a highly aggressive malignancy with an extremely poor prognosis, whereas pancreatic neuroendocrine tumors (PNETs) are relatively rare and generally less aggressive pancreatic neoplasms (Stoop et al 2025; Bray et al 2024; Chauhan et al 2024; Rikhraj et al 2025). However, a subset of PNETs may show atypical hypoenhancement on contrast-enhanced CT because of abundant fibrous stroma, thereby mimicking PDAC and increasing the risk of misdiagnosis (Ronot et al 2017; Jeon et al 2017). Given the substantial differences between PDAC and PNETs in biological behavior, prognosis, and treatment strategy, accurate preoperative differentiation is of considerable clinical importance (Jiang et al 2025; Bengtsson et al 2020; Daamen et al 2022).
Although previous CT radiomics studies have explored the differentiation between pancreatic neuroendocrine tumors and PDAC, the remaining gap lies in the limited evaluation of hypo-PNETs specifically, the insufficient comparison between single-phase and multiphase imaging strategies, and the lack of systematic assessment of whether clinical integration provides meaningful added value beyond radiomics alone (He et al 2019 ; Yang et al 2019 ; Qureshi et al 2022). This issue is particularly important for hypo-PNETs, because unlike typical PNETs, they often show atypical hypoenhancement and are therefore much more likely to be mistaken for PDAC on routine CT. The inclusion of the non-contrast phase may provide complementary information beyond enhanced imaging alone, because baseline attenuation, internal composition, necrosis, hemorrhage, and calcification can be better appreciated before contrast administration. These intrinsic tumor characteristics may supplement the vascular information obtained from arterial- and venous-phase imaging and thereby improve lesion characterization. Therefore, the present study was designed to address three related questions in the differential diagnosis between hypo-PNETs and PDAC: first, whether multiphase radiomics provides better diagnostic information than single-phase radiomics; second, which machine-learning classifier performs best for this task; and third, whether integrating the optimal radiomics signature with clinical factors provides incremental value over either approach alone.
Materials and methods
Study population
This retrospective study enrolled patients with pathologically confirmed PNETs and PDACs treated at Jiangsu Province Hospital of Chinese Medicine, Affiliated Hospital of Nanjing University of Chinese Medicine, from January 2012 to June 2025. Inclusion criteria were: (1) preoperative imaging performed within 2 weeks before surgery; (2) postoperative pathological confirmation of tumor type; and (3) complete clinical and imaging data with sufficient image quality. Exclusion criteria were: (1) prior treatment before imaging, including chemotherapy or radiotherapy; (2) poor-quality CT images or availability of only MR images; (3) PNET lesions with arterial phase enhancement greater than that of the surrounding pancreatic parenchyma; and (4) recurrent disease or the presence of other tumors. Our institutional review board approved this retrospective study (Approval No.2022NL-060-02). The requirement for written informed consent was waived due to the retrospective nature of the study.
CT examination protocol
CT examinations were performed using multidetector CT scanners, including the GE Discovery HD750, GE Optima 670, and Philips Brilliance 64. Although minor differences in hardware platform existed among scanners, the examinations were performed using a comparable pancreatic CT protocol at our institution. The major acquisition parameters were kept as consistent as possible across scanners, including a tube voltage of 120 kVp, tube current of 120–400 mAs with automatic adjustment when applicable, pitch of 1.375, slice thickness of 3.0 mm, slice interval of 3.0 mm, and reconstruction thickness of 1.25 mm. To further reduce inter-scanner variability before radiomics analysis, all images were resampled to isotropic voxels and underwent unified preprocessing and intensity standardization.
For contrast-enhanced CT, 40 mL of saline was first injected through the median cubital vein using a power injector, followed by administration of a nonionic contrast agent (Ultravist 300, 1.2 mL/kg) at an injection rate of 3.0 mL/s. Biphasic contrast-enhanced images were acquired at 30 s for the arterial phase and 60 s for the portal venous phase after contrast administration.
Clinical and imaging data analysis
Patient imaging and clinical data were obtained from the routine clinical records and the Picture Archiving and Communication System (PACS) of our institution. The retrospectively collected clinical variables included age, sex, and serum carbohydrate antigen 19–9 (CA19-9) levels. All CT images were independently reviewed by two radiologists with 3 and 7 years of experience in abdominal CT interpretation, respectively, and any discrepancies were resolved by consensus. Both readers were blinded to the histopathological findings and clinical data. The evaluated imaging features included maximum tumor diameter, tumor texture, tumor location, lesion margin, pancreatic duct dilatation, bile duct dilatation, pancreatic atrophy, associated pancreatitis, calcification, and lymph node enlargement.
CT image analysis
Image preprocessing and segmentation
Before feature extraction, all CT images were resampled to an isotropic voxel size of 1 mm × 1 mm × 1 mm to improve feature robustness and reproducibility. Intensity standardization was then performed using a fixed abdominal soft-tissue window setting (window width, 350 HU; window level, 50 HU) before manual segmentation.
Non-contrast, arterial-phase, and venous-phase images were exported from the PACS in DICOM format and imported into ITK-SNAP software. Whole-tumor volumes of interest (VOIs) were manually delineated slice by slice along the lesion margins on each phase separately by two radiologists with 3 and 7 years of experience in abdominal imaging, respectively. During segmentation, intratumoral necrotic and calcified areas were included, whereas the surrounding normal pancreatic parenchyma and adjacent large vessels were carefully excluded.
To ensure inter-phase consistency, the readers performed segmentation with reference to the corresponding lesion location, tumor contour, maximum diameter, and adjacent anatomical landmarks across the non-contrast, arterial-phase, and venous-phase images. When the tumor boundary was ambiguous in a given phase, the other phases were reviewed simultaneously to improve boundary matching and maintain anatomical consistency. In cases of disagreement, the segmentation boundaries were reviewed and finalized by a senior radiologist with 10 years of experience in abdominal imaging. The resulting VOI data were saved in NRRD format (Fig. 1).
Fig. 1.
Workflow of radiomics analysis for differentiating hypo-PNET from PDAC
The radiomics workflow consisted of four steps: VOI segmentation, feature extraction, feature selection, and model development and evaluation. First, tumor regions were manually delineated on CT images in axial, coronal, and sagittal planes. Second, radiomics features were extracted from the segmented volumes of interest. Third, feature selection was performed through a multistep procedure, including preliminary statistical screening, Spearman correlation analysis, minimum redundancy maximum relevance (mRMR) analysis, and least absolute shrinkage and selection operator (LASSO) regression. Finally, predictive models were constructed and evaluated using receiver operating characteristic (ROC) analysis, decision curve analysis (DCA), and calibration curves.
Feature extraction and selection
Radiomics features were extracted from the segmented volumes of interest using Python 3.7 and the PyRadiomics package. The extracted handcrafted features were divided into three categories: geometry, intensity, and texture features. Geometry features describe the three-dimensional shape characteristics of the tumor. Intensity features describe the first-order statistical distribution of voxel intensities within the tumor. Texture features describe second- and higher-order spatial distributions of voxel intensities and were derived using the gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM) methods.
To evaluate feature reproducibility, 30 cases were randomly selected for interobserver and intraobserver analyses. Two radiologists independently delineated the VOIs, and one radiologist repeated the segmentation after a 2-week interval. Intraclass correlation coefficients (ICCs) were calculated for the extracted radiomics features to assess interobserver and intraobserver agreement. Features with both interobserver and intraobserver ICC values > 0.75 were considered reproducible and were retained for subsequent analysis.
Feature selection was performed in four steps using the training set only. First, radiomics features with P < 0.05 in the univariate statistical analysis were retained. Second, Spearman’s rank correlation analysis was used to remove highly correlated features; when the correlation coefficient between two features exceeded 0.9, one of the features was excluded. Third, minimum redundancy maximum relevance (mRMR) analysis was applied to further reduce feature redundancy and retain the most informative features. Finally, the least absolute shrinkage and selection operator (LASSO) regression model was applied for further feature selection. The penalty parameter (λ) was determined using tenfold cross-validation with the minimum-criteria rule, and features with nonzero coefficients were retained to construct the radiomics signature.
Machine learning classifier construction
After LASSO-based feature selection, the retained radiomics features from the non-contrast, arterial-phase, venous-phase, and combined three-phase images were entered into multiple machine-learning classifiers, including logistic regression (LR), support vector machine (SVM), random forest (RF), naive Bayes (NB), and k-nearest neighbors (KNN), to compare radiomics model performance across different algorithms. All model development procedures were performed using the training set only. Because this was a retrospective exploratory study with a relatively limited sample size, the classifiers were implemented using predefined settings rather than an extensive nested hyperparameter optimization procedure. Five-fold cross-validation was performed within the training set for internal model development and model comparison. The classifier with the highest mean cross-validated AUC in the radiomics analysis was selected as the unified modeling algorithm for subsequent construction of the clinical, radiomics, and combined models.
Construction and validation of the clinical and combined models
Univariate logistic regression analysis was first performed in the training set for all candidate clinical and imaging variables. Variables with P < 0.05 were subsequently entered into multivariate logistic regression analysis to identify independent clinical predictors. After the optimal classifier had been determined from the radiomics analysis, these independent clinical predictors were used to construct the clinical model within the same classification framework. Subsequently, the combined model was established by integrating the selected clinical predictors with the optimal radiomics features, allowing direct comparison among the clinical, radiomics, and combined models.
Statistical analysis
Statistical analysis was performed using SPSS 27.0 (IBM Corp., Armonk, NY, USA) and Python 3.7. Continuous variables were first assessed for normality. Normally distributed variables are presented as mean ± standard deviation and were compared using the independent-samples t test, whereas non-normally distributed variables are presented as median (interquartile range) and were compared using the Mann–Whitney U test. Categorical variables are presented as counts and percentages and were compared using the chi-square test or Fisher’s exact test, as appropriate.
The entire cohort was randomly divided into a training set and a test set at a ratio of 7:3 using stratified sampling. This split was performed once, and the test set was kept fixed throughout the analysis to ensure fair comparison across different models. All preprocessing, feature selection, and model development procedures were performed using the training set only, whereas the test set was used solely for final evaluation.
For radiomics feature selection, features with P < 0.05 in the univariate statistical analysis were first retained. Highly correlated features were then removed using Spearman’s rank correlation analysis. Subsequently, minimum redundancy maximum relevance (mRMR) analysis and least absolute shrinkage and selection operator (LASSO) regression were applied for further feature selection. The penalty parameter (λ) in the LASSO model was determined using tenfold cross-validation with the minimum-criteria rule, and features with nonzero coefficients were retained to construct the radiomics signature.
Clinical and imaging variables with P < 0.05 in univariate analysis were entered into multivariate logistic regression analysis to identify independent clinical predictors. Radiomics models were initially developed using multiple machine-learning classifiers, including LR, SVM, RF, NB, and KNN. Five-fold cross-validation was performed within the training set for internal model development and model comparison. The classifier with the highest mean cross-validated AUC in the radiomics analysis was then selected as the unified modeling algorithm for subsequent construction of the clinical, radiomics, and combined models. The clinical model was built using the selected independent clinical predictors, and the combined model was established by integrating the optimal radiomics features with these clinical predictors within the same classification framework.
The diagnostic performance of each model was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Pairwise comparisons of AUCs were performed using the DeLong test to determine whether differences in discrimination between models were statistically significant. In addition, net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were calculated to further assess whether the combined model provided incremental value beyond the clinical-only or radiomics-only model. Specifically, NRI was used to quantify improvement in case reclassification, whereas IDI was used to evaluate the overall gain in discrimination. These analyses were performed separately in the training set and test set. Calibration curves and the Hosmer–Lemeshow test were used to assess the agreement between predicted and observed probabilities, and decision curve analysis (DCA) was performed to evaluate clinical utility. A two-sided P < 0.05 was considered statistically significant.
Results
Analysis of clinical factors
A total of 297 patients were included in this study, comprising 99 patients with hypo-PNET and 198 patients with PDAC. The cohort was randomly stratified into a training set and a test set at a ratio of 7:3. The clinical characteristics and CT imaging features of the enrolled patients are summarized in Table 1. Clinical variables and CT imaging features were entered into univariate and multivariate logistic regression analyses to identify independent predictors for differentiating hypo-PNET from PDAC. The results showed that pancreatic duct dilatation, tumor composition, age, and maximum tumor diameter were independent predictors for distinguishing hypo-PNET from PDAC (all P < 0.05; Table 2).
Table 1.
Clinical and imaging characteristics of patients in the training and test sets
| Characteristic | Training set (n = 207) | Test set (n = 90) | ||||
|---|---|---|---|---|---|---|
| PDACs (n = 140) | hypo-PNETs (n = 67) | P value | PDACs (n = 58) | hypo-PNETs (n = 32) | P value | |
| Age (years) | 63.42 ± 9.01 | 54.96 ± 10.98 | < 0.001 | 64.50 ± 9.29 | 57.24 ± 12.22 | 0.003 |
| Maximum tumor diameter (cm) | 3.47 ± 1.38 | 3.84 ± 1.94 | 0.516s | 3.44 ± 1.61 | 4.02 ± 1.94 | 0.136 |
| Sex | 0.016 | < 0.001 | ||||
| Female | 55(39.29) | 39(58.21) | 13(22.41) | 20(62.50) | ||
| Male | 85(60.71) | 28(41.79) | 45(77.59) | 12(37.50) | ||
| CA19-9 n(%) | 0.053 | 0.01 | ||||
| Normal | 64(45.71) | 41(61.19) | 22(37.93) | 22(68.75) | ||
| Abnormal | 76(54.29) | 26(38.81) | 36(62.07) | 10(31.25) | ||
| Tumor composition n(%) | < 0.001 | < 0.001 | ||||
| Cystic-solid | 15(10.71) | 37(55.22) | 5(8.62) | 19(59.38) | ||
| solid | 125(89.29) | 30(44.78) | 53(91.38) | 13(40.62) | ||
| Tumor location n(%) | 0.009 | 0.527 | ||||
| Head/neck | 61(43.57) | 43(64.18) | 29(50.00) | 19(59.38) | ||
| Body/tail | 79(56.43) | 24(35.82) | 29(50.00) | 13(40.62) | ||
| Margin n(%) | 0.007 | 0.016 | ||||
| Well-defined | 6(4.29) | 11(16.42) | 3(5.17) | 8(25.00) | ||
| Ill-defined | 134(95.71) | 56(83.58) | 55(94.83) | 24(75.00) | ||
| Pancreatic duct dilatation n(%) | < 0.001 | 0.009 | ||||
| Absent | 36(25.71) | 50(74.63) | 20(34.48) | 21(65.62) | ||
| Present | 104(74.29) | 17(25.37) | 38(65.52) | 11(34.38) | ||
| Bile duct dilatation n(%) | < 0.001 | 0.926 | ||||
| Absent | 75(53.57) | 53(79.10) | 31(53.45) | 16(50.00) | ||
| Present | 65(46.43) | 14(20.90) | 27(46.55) | 16(50.00) | ||
| Pancreatic atrophy n(%) | 0.003 | 0.463 | ||||
| Absent | 87(62.14) | 56(83.58) | 35(57.38) | 26(78.79) | ||
| Present | 53(37.86) | 11(16.42) | 26(42.62) | 7(21.21) | ||
| Associated pancreatitis n(%) | 0.028 | 1.000 | ||||
| Absent | 114(81.43) | 63(94.03) | 51(87.93) | 27(87.50) | ||
| Present | 26(18.57) | 4(5.97) | 7(12.07) | 4(12.50) | ||
| Calcification n(%) | 0.031 | 0.692 | ||||
| Absent | 133(95.00) | 57(85.07) | 52(89.66) | 27(84.38) | ||
| Present | 7(5.00) | 10(14.93) | 6(10.34) | 5(15.62) | ||
| Lymph node enlargement n(%) | 0.005 | 0.003 | ||||
| Absent | 72(51.43) | 49(73.13) | 27(46.55) | 26(81.25) | ||
| Present | 68(48.57) | 18(26.87) | 31(53.45) | 6(18.75) | ||
Table 2.
Univariate and multivariate logistic regression analyses of clinical and imaging characteristics
| Variable | Univariate analysis | Multivariate analysis | ||
|---|---|---|---|---|
| OR (95%CI) | P value | OR (95%CI) | P value | |
| Associated pancreatitis | 0.154 (0.054–0.372) | < 0.001 | 0.321 (0.109–0.938) | 0.081 |
| Pancreatic duct dilatation | 0.163 (0.106–0.251) | < 0.001 | 0.182 (0.091–0.366) | < 0.001 |
| Tumor composition | 0.240 (0.172–0.335) | < 0.001 | 0.203 (0.101–0.408) | < 0.001 |
| Lymph node enlargement | 0.265 (0.171–0.409) | < 0.001 | 0.570 (0.290–1.119) | 0.170 |
| Sex | 0.329 (0.230–0.471) | < 0.001 | 0.682 (0.356–1.306) | 0.333 |
| CA19-9 | 0.342 (0.236–0.497) | < 0.001 | 0.428 (0.170–1.074) | 0.072 |
| Tumor location | 0.304 (0.207–0.446) | < 0.001 | 0.622 (0.327–1.184) | 0.225 |
| Bile duct dilatation | 0.215 (0.133–0.350) | < 0.001 | 0.729 (0.348–1.528) | 0.483 |
| Pancreatic atrophy | 0.208 (0.120–0.358) | < 0.001 | 0.501 (0.228–1.104) | 0.150 |
| Maximum tumor diameter | 0.860 (0.807–0.916) | < 0.001 | 1.288 (1.060–1.565) | 0.033 |
| Age | 0.986 (0.982–0.990) | < 0.001 | 1.026 (1.007–1.046) | 0.023 |
| Calcification | 1.429 (0.635–3.212) | 0.469 | ||
| Margin | 1.833 (0.795–4.225) | 0.232 | ||
Radiomics feature selection
A total of 1,834 radiomics features were extracted from each of the non-contrast, arterial-phase, and venous-phase CT images, yielding 5,502 features in the combined three-phase dataset. After multistep feature selection in the training set, 13, 17, 11, and 15 optimal radiomics features were retained from the non-contrast, arterial-phase, venous-phase, and combined three-phase images, respectively.
Construction of radiomics models
Radiomics models were constructed using five machine-learning classifiers, namely LR, RF, SVM, NB, and KNN, based on the selected features from the non-contrast, arterial-phase, venous-phase, and combined three-phase images. Their classification performance was evaluated using ROC analysis (Fig. 2), and the detailed performance metrics are provided in Online Resource 2. The combined three-phase model outperformed the models based on a single phase. Among the five classifiers, the SVM-based model demonstrated the best predictive performance and was therefore selected as the optimal radiomics model, with AUCs of 0.814 and 0.812 in the training and test sets, respectively.
Fig. 2.
ROC curves of radiomics models constructed using different machine-learning classifiers based on different CT image sets
Panels A and B show the ROC curves of radiomics models based on non-contrast CT in the training and test sets, respectively. Panels C and D show the corresponding results for arterial-phase CT. Panels E and F show the corresponding results for venous-phase CT. Panels G and H show the corresponding results for the combined three-phase CT image set. Models were developed using five machine-learning classifiers, including LR, SVM, RF, NB, and KNN.
Construction and validation of the combined model
Based on logistic regression analysis, pancreatic duct dilatation, tumor composition, age, and maximum tumor diameter were identified as independent clinical predictors. In the radiomics classifier comparison, SVM demonstrated the best performance and was therefore selected as the unified classifier for subsequent model construction. Using this classifier, the clinical model, radiomics model, and combined model were constructed and compared. Numerically, the combined model achieved the highest AUCs in both the training and test cohorts (Table 3), reaching 0.886 (95% CI 0.851–0.931) and 0.849 (95% CI 0.746–0.931), respectively. DCA curves also suggested that the combined model may provide higher net benefit across a wider range of threshold probabilities, and the calibration curves demonstrated acceptable agreement between predicted and observed outcomes. However, because the incremental gain over the radiomics-only model in the test cohort was limited and the corresponding between-model comparisons did not reach statistical significance, this advantage should be interpreted cautiously (Fig. 3).
Table 3.
Diagnostic performance of the clinical, radiomics, and combined models in the training and test sets
| AUC (95%CI) | Accuracy | Sensitivity | Specificity | PPV | NPV | ||
|---|---|---|---|---|---|---|---|
| Radiomics | Training set | 0.814 (0.751–0.878) | 0.778 | 0.731 | 0.800 | 0.636 | 0.862 |
| Test set | 0.812 (0.720–0.905) | 0.722 | 0.844 | 0.679 | 0.563 | 0.913 | |
| Clinical | Training set | 0.801 (0.736–0.866) | 0.739 | 0.866 | 0.729 | 0.620 | 0.953 |
| Test set | 0.755 (0.646–0.865) | 0.733 | 0.781 | 0.655 | 0.574 | 0.884 | |
| Combined | Training set | 0.886 (0.841–0.931) | 0.792 | 0.792 | 0.707 | 0.595 | 0.854 |
| Test set | 0.849 (0.767–0.931) | 0.767 | 0.767 | 0.707 | 0.622 | 0.911 |
Fig. 3.
ROC, decision curve, and calibration analyses of the radiomics, clinical, and combined models
The diagnostic performance and clinical utility of the radiomics model (A_V_N), clinical model, and combined model were evaluated in both cohorts. Panels A and D show the ROC curves of the three models in the training and test sets, respectively. Panels B and E show the decision curve analyses, demonstrating the net benefit of each model across different threshold probabilities. Panels C and F show the calibration curves of the three models in the training and test sets, respectively. The dashed diagonal line indicates perfect calibration; curves closer to this line represent better agreement between predicted and observed probabilities.
Comparison of model performance and incremental value
Pairwise DeLong tests as well as NRI and IDI analyses were further performed to compare the discriminative and incremental value of the three models, and the detailed results are presented in Online Resource 3. In the training cohort, the combined model showed significantly better performance than the clinical model (P = 0.001), whereas no significant difference was found between the radiomics and clinical models (P = 0.770) or between the combined and radiomics models (P = 0.070). In the test cohort, none of the pairwise comparisons reached statistical significance, including radiomics versus clinical (P = 0.459), combined versus radiomics (P = 0.550), and combined versus clinical (P = 0.051).
The incremental value of the combined model was further evaluated using NRI and IDI (Online Resource 3). For NRI, the combined model yielded values of 0.123 and 0.110 over the radiomics and clinical models, respectively, in the training cohort, and 0.083 and 0.094, respectively, in the test cohort. For IDI, the corresponding improvements were 0.182 and 0.280 in the training cohort and 0.134 and 0.247 in the test cohort.
Significance testing for IDI showed that, in the training cohort, the combined model significantly improved discrimination compared with the clinical model (P = 0.004), whereas the comparison between the combined and radiomics models did not reach significance (P = 0.066). No significant difference was observed between the radiomics and clinical models (P = 0.337). In the test cohort, none of the IDI comparisons were statistically significant, including combined versus radiomics (P = 0.363), combined versus clinical (P = 0.092), and radiomics versus clinical (P = 0.452).
Overall, the combined model showed numerically improved discrimination and reclassification compared with the single-modality models; however, this advantage was more evident in the training cohort, whereas in the test cohort the between-model differences did not reach statistical significance.
Discussion
Patients with hypo-PNET and PDAC have markedly different prognoses. Therefore, accurate preoperative differentiation between these two entities is crucial for individualized treatment planning. In the present study, we developed a predictive model for differentiating hypo-PNET from PDAC based on multiphase contrast-enhanced CT radiomics features in combination with important clinical factors. Our results showed that the combined three-phase radiomics model achieved higher diagnostic performance than the single-phase radiomics models. In addition, the combined model integrating clinical factors with three-phase radiomics features yielded the highest AUCs numerically in both the training and test cohorts. However, because several between-model comparisons in the test cohort did not reach statistical significance, these findings should be interpreted as indicating a potential incremental value of the combined model rather than definitive superiority.
CA19-9 was significant in the univariate analysis but did not remain an independent predictor in the multivariate model. A possible explanation is that CA19-9 is influenced not only by tumor biology but also by biliary obstruction, inflammatory changes, and overall tumor burden, which may reduce its specificity in this differential diagnosis. In addition, some of its discriminatory information may overlap with imaging-derived variables, thereby weakening its independent effect after multivariable adjustment.
Some variables, such as bile duct dilatation and pancreatic atrophy, showed statistical significance in the training set but not in the test set. This inconsistency may be related to the smaller sample size of the test set and the sampling fluctuation introduced by a single random split. Therefore, these discrepancies should be interpreted cautiously and should not be overemphasized.
Although only a limited number of optimal radiomics features were ultimately retained for each phase, these features were predominantly texture-related features, together with a smaller proportion of first-order and shape descriptors. This finding suggests that the differential diagnosis between hypo-PNETs and PDAC depended mainly on differences in intratumoral heterogeneity and enhancement-related spatial complexity rather than on size alone. Texture features may reflect underlying differences in stromal composition, fibrosis, necrosis, and microvascular architecture, whereas shape-related features may be associated with distinct growth patterns, such as the more infiltrative tendency of PDAC and the relatively expansile growth pattern of some hypo-PNETs. In addition, the superior performance of the combined three-phase model indicates that multiphasic imaging may provide complementary biological information that cannot be fully captured by any single phase alone.
PDAC and PNETs differ substantially in histological composition, vascular characteristics, and tumor-related stromal response, which provides the biological basis for imaging-based differentiation (Karmazanovsky et al 2019). PDAC is typically characterized by abundant desmoplastic stroma, glandular destruction, and low microvascular density. As a result, PDAC usually exhibits hypoenhancement in both the arterial and venous phases and is often accompanied by main pancreatic duct dilatation, focal pancreatic atrophy, and bile duct dilatation (Pratticò et al., 2024). In the present study, logistic regression analysis identified pancreatic duct dilatation, tumor composition, age and maximum tumor diameter as independent predictors, which is consistent with the abovementioned pathological characteristics. PNETs are generally hypervascular tumors; however, some hypo-PNETs show atypical enhancement because of an increased proportion of fibrous tissue, resulting in enhancement patterns on contrast-enhanced CT that resemble those of PDAC (Konukiewitz et al 2022). Nevertheless, hypo-PNETs may still retain a certain degree of microvascular architecture and relatively homogeneous tissue morphology. Therefore, their radiomics features, including texture, tumor composition, and shape descriptors, may still differ from those of PDAC. The finding that the combined three-phase radiomics model outperformed the single-phase models further suggests that tissue architecture and dynamic vascular changes captured by multiphase contrast-enhanced CT are critical for distinguishing these two tumor types.
Radiomics has shown promising value in the differential diagnosis of pancreatic tumors by enabling quantitative characterization of tumor heterogeneity (Podină et al 2025; Mayerhopefer et al 2020; Rogers et al 2020). Previous studies have demonstrated its potential to improve diagnostic accuracy in distinguishing PNET from PDAC. For instance, the study (Zhuang et al 2025) developed radiomics models based on arterial- and venous-phase CT images and found that the random forest model with quantile transformation achieved the best diagnostic performance. Our findings are generally consistent with theirs, but the present study offers several additional strengths. Specifically, we incorporated non-contrast CT into the analysis and demonstrated that combined three-phase features outperformed single-phase models, highlighting the added value of multiphasic imaging. We also systematically compared five machine-learning classifiers and found that SVM yielded the best performance, which may be related to its advantages in handling high-dimensional and nonlinear data (Anai et al 2022). Furthermore, the integration of clinical factors with radiomics features improved the generalizability and interpretability of the model. Collectively, these findings further support the clinical value of radiomics for pancreatic tumor differentiation and provide practical insight into the use of multiphase CT and appropriate machine learning strategies.
PDAC and PNET differ substantially in clinical management strategies and prognosis. PDAC is associated with poor prognosis and aggressive biological behavior, and it usually requires multimodal treatment. In contrast, PNETs, particularly localized and low-grade lesions, may achieve favorable outcomes after surgery or local therapy. In clinical practice, some hypo-PNETs exhibit atypical enhancement patterns that overlap with those of PDAC, which may lead to misdiagnosis. The combined model developed in this study provides an objective and quantitative tool for diagnostic support. When conventional CT findings are insufficient for definitive differentiation, this model may offer complementary information for preoperative assessment and multidisciplinary team (MDT) decision-making, thereby potentially helping reduce misdiagnosis and overtreatment.
Several limitations of this study should be acknowledged. First, this was a retrospective study using images acquired from multiple CT scanners. Although standardized preprocessing was performed, residual scanner-related variability could not be completely eliminated. Future multicenter studies with more harmonized acquisition protocols and image harmonization strategies are needed. Second, only internal validation was performed. Although the test set provided preliminary validation, the lack of external validation limits generalizability. Third, the dataset was divided using a single random train–test split rather than repeated resampling or nested cross-validation, which may result in unstable performance estimates in a relatively limited cohort. Fourth, manual segmentation is labor-intensive and may limit large-scale implementation; future studies should explore semi-automatic or fully automatic segmentation methods.
From a translational perspective, the combined model may serve as a decision-support tool rather than a standalone diagnostic system. In future work, it could be converted into a nomogram or a web-based calculator to facilitate clinical application. Because contrast-enhanced CT is routinely available, such a model may be particularly useful in centers with limited expertise in pancreatic tumor imaging, where it could provide complementary information for radiologists and multidisciplinary teams when conventional imaging findings overlap.
Conclusions
In conclusion, multiphasic CT radiomics, particularly the combined three-phase radiomics signature, showed value in differentiating hypo-PNETs from PDAC. Integrating clinical factors further improved overall model performance numerically; however, because the incremental value in the test set was limited, the superiority of the combined model should be interpreted cautiously. External validation in larger multicenter cohorts is warranted before clinical implementation.
Supplementary Information
Below is the link to the electronic supplementary material.
Author contributions
All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Fan Xia, Jie Yu, Jianhua Wang, and Zhongqiu Wang. The first draft of the manuscript was written by Fan Xia, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.
Funding
This work was supported by the Jiangsu Provincial Key Research and Development Program (Grant No. BE2023789) and the National Natural Science Foundation of China (Grant Nos. 82371919 and 82171925).
Data availability
The datasets generated during and/or analysed during the current study are not publicly available because they are derived from clinical imaging records and associated clinical information involving patient privacy, but are available from the corresponding author on reasonable request and subject to institutional approval.
Declarations
Competing interests
The authors declare no competing interests.
Ethics approval
This retrospective study was approved by the Institutional Review Board of the Affiliated Hospital of Nanjing University of Chinese Medicine (Approval No. 2022NL-060–02) and was performed in accordance with the principles of the Declaration of Helsinki.
Consent to participate
The requirement for written informed consent was waived by the Institutional Review Board because of the retrospective nature of the study.
Consent for publication
Not applicable.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Contributor Information
Jianhua Wang, Email: wangjianhua84@163.com.
Zhongqiu Wang, Email: zhongqiuwang@njucm.edu.cn.
References
- Anai K, Hayashida Y, Ueda I, Hozuki E, Yoshimatsu Y, Tsukamoto J, Hamamura T, Onari N, Aoki T, Korogi Y (2022) The effect of CT texture-based analysis using machine learning approaches on radiologists’ performance in differentiating focal-type autoimmune pancreatitis and pancreatic duct carcinoma. Jpn J Radiol 40:1156–1165. 10.1007/s11604-022-01298-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Bengtsson A, Andersson R, Ansari D (2020) The actual 5-year survivors of pancreatic ductal adenocarcinoma based on real-world data. Sci Rep 10:16425. 10.1038/s41598-020-73525-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 74:229–263. 10.3322/caac.21834 [DOI] [PubMed] [Google Scholar]
- Chauhan A, Chan K, Halfdanarson TR, Bellizzi AM, Rindi G, O’Toole D, Ge PS, Jain D, Dasari A, Anaya DA, Bergsland E, Mittra E, Wei AC, Hope TA, Kendi AT, Thomas SM, Flem S, Brierley J, Asare EA, Washington K, Shi C (2024) Critical updates in neuroendocrine tumors: version 9 American Joint Committee on Cancer staging system for gastroenteropancreatic neuroendocrine tumors. CA Cancer J Clin 74:359–367. 10.3322/caac.21840 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Daamen LA, Dorland G, Brada LJH, Groot VP, van Oosten AF, Besselink MG, Bosscha K, Bonsing BA, Busch OR, Cirkel GA, van Dam RM, Festen S, Groot Koerkamp B, Haj Mohammad N, van der Harst E, de Hingh IHJT, Intven MPW, Kazemier G, Los M, de Meijer VE, Nieuwenhuijs VB, Roos D, Schreinemakers JMJ, Stommel MWJ, Verdonk RC, Verkooijen HM, Molenaar IQ, van Santvoort HC, Dutch Pancreatic Cancer Group (2022) Preoperative predictors for early and very early disease recurrence in patients undergoing resection of pancreatic ductal adenocarcinoma. HPB (Oxford) 24:535–546. 10.1016/j.hpb.2021.09.004 [DOI] [PubMed] [Google Scholar]
- He M, Liu Z, Lin Y, Wan J, Li J, Xu K, Wang Y, Jin Z, Tian J, Xue H (2019) Differentiation of atypical non-functional pancreatic neuroendocrine tumor and pancreatic ductal adenocarcinoma using CT based radiomics. Eur J Radiol 117:102–111. 10.1016/j.ejrad.2019.05.024 [DOI] [PubMed] [Google Scholar]
- Jeon SK, Lee JM, Joo I, Lee ES, Park HJ, Jang JY, Ryu JK, Lee KB, Han JK (2017) Nonhypervascular pancreatic neuroendocrine tumors: differential diagnosis from pancreatic ductal adenocarcinomas at MR imaging-retrospective cross-sectional study. Radiology 284:77–87. 10.1148/radiol.2016160586 [DOI] [PubMed] [Google Scholar]
- Jiang XY, Jiang HT, Liu YH (2025) Research progress in the treatment of pancreatic neuroendocrine neoplasms. Chin J Pancreatol 25:312–316. 10.3760/cma.j.cn115667-20241030-00180 [Google Scholar]
- Karmazanovsky G, Belousova E, Schima W, Glotov A, Kalinin D, Kriger A (2019) Nonhypervascular pancreatic neuroendocrine tumors: spectrum of MDCT imaging findings and differentiation from pancreatic ductal adenocarcinoma. Eur J Radiol 110:66–73. 10.1016/j.ejrad.2018.04.006 [DOI] [PubMed] [Google Scholar]
- Konukiewitz B, Jesinghaus M, Kasajima A, Klöppel G (2022) Neuroendocrine neoplasms of the pancreas: diagnosis and pitfalls. Virchows Arch 480:247–257. 10.1007/s00428-021-03211-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Mayerhoefer ME, Materka A, Langs G, Häggström I, Szczypiński P, Gibbs P, Cook G (2020) Introduction to radiomics. J Nucl Med 61:488–495. 10.2967/jnumed.118.222893 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Podină N, Gheorghe EC, Constantin A, Cazacu I, Croitoru V, Gheorghe C, Balaban DV, Jinga M, Țieranu CG, Săftoiu A (2025) Artificial intelligence in pancreatic imaging: a systematic review. Gastroenterol J 13:55–77. 10.1002/ueg2.12723 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Pratticò F, Garajová I (2024) Focus on pancreatic cancer microenvironment. Curr Oncol 31:4241–4260. 10.3390/curroncol31080316 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Qureshi TA, Gaddam S, Wachsman AM, Wang L, Azab L, Asadpour V, Chen W, Xie Y, Wu B, Pandol SJ, Li D (2022) Predicting pancreatic ductal adenocarcinoma using artificial intelligence analysis of pre-diagnostic computed tomography images. Cancer Biomark 33:211–217. 10.3233/CBM-210273 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Rikhraj N, Fernandez CJ, Ganakumar V, Pappachan JM (2025) Pancreatic neuroendocrine tumors: a case-based evidence review. World J Gastrointest Pathophysiol 16:107265. 10.4291/wjgp.v16.i2.107265 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Rogers W, Thulasi Seetha S, Refaee TAG, Lieverse RIY, Granzier RWY, Ibrahim A, Keek SA, Sanduleanu S, Primakov SP, Beuque MPL, Marcus D, van der Wiel AMA, Zerka F, Oberije CJG, van Timmeren JE, Woodruff HC, Lambin P (2020) Radiomics: from qualitative to quantitative imaging. Br J Radiol 93:20190948. 10.1259/bjr.20190948 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ronot M, Cuccioli F, Dioguardi Burgio M, Vullierme MP, Hentic O, Ruszniewski P, d’Assignies G, Vilgrain V (2017) Neuroendocrine liver metastases: vascular patterns on triple-phase MDCT are indicative of primary tumour location. Eur J Radiol 89:156–162. 10.1016/j.ejrad.2017.02.007 [DOI] [PubMed] [Google Scholar]
- Stoop TF, Javed AA, Oba A, Koerkamp BG, Seufferlein T, Wilmink JW, Besselink MG (2025) Pancreatic cancer. Lancet 405:1182–1202. 10.1016/S0140-6736(25)00261-2 [DOI] [PubMed] [Google Scholar]
- Yang CW, Jiang HY, Liu XJ, Song B (2019) Research progress of radiomics in the imaging evaluation of pancreatic tumor lesions. Radiol Pract 34:963–968. 10.13609/j.cnki.1000-0313.2019.09.006 [Google Scholar]
- Zhuang Y, Cui JJ, Yang GJ, Li B, Wang N, Sun HK, Wang ZG (2025) The value of radiomic models based on contrast-enhanced computed tomography images in distinguishing between pancreatic ductal adenocarcinoma and pancreatic neuroendocrine neoplasms. J Precis Med 40:265–269. 10.13362/j.jpmed.202540071 [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets generated during and/or analysed during the current study are not publicly available because they are derived from clinical imaging records and associated clinical information involving patient privacy, but are available from the corresponding author on reasonable request and subject to institutional approval.