Development and validation of a clinical-radiomics nomogram for the differential diagnosis of focal pancreatic solid lesions: a retrospective cohort study

Introduction

Focal pancreatic solid lesions (FPSLs) represent a spectrum of pathological conditions that appear as localized masses within the pancreas (1). Focal autoimmune pancreatitis (fAIP) and mass-forming chronic pancreatitis (MFCP) are the most common benign conditions, accounting for approximately 10–30% and 28–41% of cases, respectively (2,3). These benign lesions generally have a favorable outlook. In contrast, the prognosis of malignant FPSLs, most notably pancreatic ductal adenocarcinoma (PDAC), has only a 5-year survival rate of less than 10% (4).

Early detection and accurate characterization of FPSLs are essential for ensuring appropriate management. This is particularly important for those benign cases, as it can protect them from the risk and complications of an unnecessary pancreaticoduodenectomy. Additionally, identifying fAIP and initiating corticosteroid treatment promptly can lead to substantial improvements in a patient’s prognosis and quality of life. However, these conditions present significant diagnostic challenges due to their often indolent and nonspecific early-stage symptoms. In fact, a considerable number of cases are detected incidentally during routine health screenings in the outpatient department, such as ultrasound, rather than through the presentation of clinical symptoms.

Subsequent diagnostic evaluations, including serum biomarker analysis and advanced imaging techniques, play a critical role in further assessing these lesions. Tumor markers like carbohydrate antigen 19-9 (CA19-9) remain a cornerstone in the clinical assessment of malignancy. However, the diagnostic utility of these markers is constrained, as their sensitivity and specificity do not meet established clinical standards. Similarly, elevated serum IgG4 levels lack complete specificity for autoimmune pancreatitis (AIP), since mild elevations can also be observed in chronic pancreatitis (CP) and in approximately 10–15% of PDAC cases (5). As a result, neither normal nor mildly abnormal CA19-9 nor IgG4 levels can reliably exclude malignancy in clinical practice.

These limitations reinforce the need for advanced imaging techniques such as contrast-enhanced computed tomography (CECT), magnetic resonance imaging (MRI), and endoscopic ultrasound (EUS) to achieve detailed characterization of FPSLs. Each technique offers unique advantages and limitations. Notably, advanced inflammatory masses can form dense, fibrotic lumps with reduced enhancement, making it difficult to distinguish from malignancy. EUS-guided fine-needle aspiration (EUS-FNA) is considered the diagnostic reference standard, with reported sensitivity and specificity of approximately 90% and 96%, respectively (6). Nonetheless, its diagnostic accuracy is highly operator-dependent and limited by sampling errors. Initial EUS-FNA can be nondiagnostic in 15–59% of cases, often due to fibrotic or paucicellular samples obtained from desmoplastic tumors or inflamed tissue (2). In summary, the significant overlap in symptoms and imaging findings, combined with the limitations of current serum markers and biopsy techniques, creates a critical need for a non-invasive method in the evaluation of FPSLs.

Artificial intelligence (AI)-based quantitative imaging, commonly referred to as “radiomics”, has emerged as a promising tool to address the limitations of conventional diagnostic methods. These data-driven imaging biomarkers provide an objective and reproducible means of characterizing lesion heterogeneity, potentially reducing interobserver variability compared to visual assessment alone. From the perspective of pathophysiology, fAIP is characterized by dense lymphoplasmacytic infiltration and relatively homogeneous “storiform” fibrosis, and MFCP primarily manifests as patchy fibrosis, acinar atrophy, and ductal distortion, whereas PDAC is associated with a prominent desmoplastic reaction and tumor cell heterogeneity. These microstructural differences are expected to manifest as quantifiable variations in computed tomography (CT) texture and heterogeneity, and can be quantified by radiomics beyond visual assessment (7-9). Meanwhile, from the perspective of quantitative imaging genomics, venous-phase CECT captures perfusion and stromal disparities of FPSLs with higher stability and reproducibility than arterial phase imaging. Clinical studies have demonstrated the potential of radiomics in differentiating PDAC from other lesions (10,11). However, most current radiomics-based research has not focused specifically on FPSLs. There is a lack of studies that simultaneously compare fAIP, MFCP, and PDAC—a triad that frequently poses a diagnostic challenge. Moreover, the potential value of specific clinical indicators beyond conventional biomarkers in distinguishing these entities has not been thoroughly investigated, and few models have been explicitly developed and validated for the outpatient setting.

In this study, a cohort of outpatients with initially suspected FPSLs who underwent CECT as part of their diagnostic workup was enrolled. The primary objective was to evaluate the diagnostic performance of a radiomics-based, integrated with key clinical predictors (such as eosinophil percentage and CA19-9) model specifically designed for FPSLs in outpatient settings. This model addresses a critical clinical need by offering a non-invasive alternative aimed at reducing the risk of preoperative misdiagnosis in outpatients with FPSLs. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-2017/rc).

Methods

Study design, patients and clinical data

This single-center retrospective cohort study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committee of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology (approval No. TJ-IRB202409044) and individual consent for this analysis was waived due to the retrospective nature. Because of the retrospective nature of case availability and the low incidence of fAIP and MFCP, a priori sample size calculation was not performed. However, all eligible patients within the study period were included.

Three groups of patients were enrolled, each representing a specific clinical diagnosis. The first group consisted of 185 AIP patients who were diagnosed according to the International Consensus Diagnostic Criteria for Autoimmune Pancreatitis between May 2013 and May 2024; The second group comprised 252 CP patients identified in specimens obtained through FNA cytology or surgery during the same period; and the third group included 451 pathologically proved PDAC patients diagnosed also by cytology or surgery between May 2022 and May 2023. Data for the patient inclusion and exclusion criteria are shown in Table S1. Finally, a total of 201 patients were enrolled in the study, which included 56 patients in the fAIP group, 51 patients in the MFCP group, and 94 patients in the PDAC group. Patients were randomly assigned to either a training set (39 fAIP patients, 35 MFCP patients, 66 PDAC patients) or a validation set (17 fAIP patients, 16 MFCP patients, 28 PDAC patients) using a 7:3 randomization ratio. Additionally, 11 patients diagnosed with fAIP and 19 patients with PDAC between June 2024 and May 2025 served as an independent temporal validation cohort. The data selection process was described in Figure 1A. For each group, comprehensive clinical variables, including demographic information, clinical presentation, and laboratory results, were retrospectively collected contemporaneously with the CECT examination (using the closest routine tests around the CT date, prior to treatment when applicable) to minimize measurement bias due to asynchrony.

Figure 1 Study flowchart and manual pancreas ROI segmentation. (A) Flow diagram showing the selection of the study population; (B) ROI of the manually segmented pancreas. AIP, autoimmune pancreatitis; CP, chronic pancreatitis; fAIP, focal autoimmune pancreatitis; MFCP, mass-forming chronic pancreatitis; PDAC, pancreatic ductal adenocarcinoma; ROI, region of interest.

Imaging segmentation, radiomics feature extraction, and model development

CT acquisition is available in Appendix 1. All CT studies were downloaded and de-identified by anonymization of Digital Imaging and Communication in Medicine (DICOM) tags, and venous phase CT images were retrieved from a picture archiving and communication system (PACS). Two radiologists who were blinded to the outcome depicted the region of interest (ROI) of the pancreas by manually outlining the pancreas independently (Figure 1B). The segmented ROIs were then exported to ITK-SNAP software for manual verification. Any incomplete outlines were manually corrected and redefined.

Radiomic feature extraction was performed using the PyRadiomics written in Python (version 3.9.20; Python Software Foundation, Wilmington, DE, USA). All extracted radiomics features were normalized using a Z-score normalization, and tested to find the best image parameter to distinguish fAIP from MFCP and PDAC. Feature selection was performed then with least absolute shrinkage and selection operator (LASSO) logistic regression algorithm. In brief, all of the features in the training group were downscaled first by t-test or the Mann-Whitney U-test to eliminate those without statistically significant differences. Multicollinearity analysis was performed to ensure the independence of selected radiomic features. To address the inherent class imbalance in the training data, we applied synthetic minority over-sampling technique (SMOTE) to augment the minority class and achieve a more balanced class distribution. LASSO with 5-fold cross-validation was then applied to select the most predictive features with non-zero coefficients, and the radiomics score (Rad-score, value of radiomics signature) was calculated as a weighted linear combination of these coefficients to construct the radiomics model.

A radiomics-clinical nomogram model was developed using multivariate logistic regression, incorporating both clinical variables and radiological features (Rad-score). Each characteristic value was assigned a score based on the regression coefficient size, and the total score was obtained by summing these individual scores. The predictive value was derived through a function that transformed the total score into the probability of the outcome event.

Statistical analysis

The statistical analysis, model creation, and performance evaluation procedures were conducted using Python (version 3.9.20), R studio (version 4.4.2, Auckland, New Zealand), and SPSS 25.0.0. Between-group comparisons of normally distributed variables were performed using the independent samples t-test, while non-normally distributed variables were compared using the Mann-Whitney U test. Categorical variables were presented as counts (percentages), and differences between groups were assessed using the chi-square test. All tests were two-sided, and a P value of <0.05 was considered statistically significant.

The diagnostic performance of each model was evaluated using receiver operating characteristic (ROC) curves, precision-recall curves (PRCs), and the area under the curve (AUC). Accuracy, sensitivity, specificity, F1 score, and other relevant evaluation metrics were calculated for each model. Decision curve analysis (DCA) was performed to quantify the net clinical benefit of each model across a continuum of threshold probabilities, comparing the models with default “treat-all” and “treat-none” strategies. These measures allow a comprehensive evaluation of both discrimination and clinical utility.

Classification thresholds were selected based on the clinical priority of each diagnostic task. To differentiate PDAC from benign lesions (fAIP or MFCP) and minimize missed malignancy, we adopted a sensitivity-first approach depending on clinical requirements: from the training cohort, we selected the threshold that achieved at least 80% sensitivity while maximizing specificity. For distinguishing between the two benign conditions (fAIP vs. MFCP), the optimal threshold was determined by maximizing the Youden index to balance sensitivity and specificity. These thresholds were applied unchanged to the validation cohorts to avoid data leakage and ensure an unbiased evaluation of generalizability.

Results

Patient characteristics, clinical features and imaging analysis

The description of the demographic characteristics, clinical symptoms and CT imaging findings is summarized in Table 1. The mean age of the patients in the three groups was 57.21±10.76, 48.25±12.14, and 60.55±9.66 years, respectively. A notable proportion of patients in all three groups were male, and a substantial number of patients were asymptomatic, with FPSL being discovered incidentally during routine physical examinations (P<0.05). The diameter of lesions in PDAC patients was the largest, and pancreatic head was the most commonly involved site across all groups (P<0.001). Imaging features also varied, with higher frequencies of pancreatic duct dilation, lymphadenopathy, and vascular involvement in PDAC patients (P<0.05). These significantly distinct imaging characteristics were consistent with the underlying pathogenesis of the diseases. For instance, fAIP more frequently involved the biliary system, whereas PDAC demonstrated a higher incidence of lymph node metastasis and vascular invasion.

Model construction and performance evaluation in discriminating fAIP from PDAC

We next analyzed the serological test data (Table S2) from the fAIP and PDAC groups. Although the globulin, CEA and CA19-9 levels showed significant differences across three subsets in the fAIP group, no statistically significant difference was observed across subsets within the PDAC cohort (Table 2). Univariate logistic regression analyses were conducted on these significant markers within the fAIP and PDAC groups in their respective training sets, and variables identified as significant in the univariate analysis were converted into dichotomous variables based on best cutoff values. Multivariate logistic regression analysis was then carried out, and revealed that eosinophil percentage (E%) and CA19-9 were significant factors distinguishing fAIP from PDAC (Table 3) (P<0.05). Given a CA19-9 level exceeding 1,000 U/mL raised clinical suspicion for pancreatic carcinoma, we classified the value of CA19-9 into three categories, that is < cutoff, between cutoff and 1,000 U/mL, and >1,000 U/mL, to construct nomogram.

A total of 851 radiomics features were then extracted from each segmentation mask, which included 18 first-order statistics of the volumetric CT intensities, 14 shape features of the target structure, 24 texture features from the gray level co-occurrence matrix (GLCM), 14 texture features from the gray level dependence matrix (GLDM), 16 texture features from the gray level size zone matrix (GLSZM), 16 texture features from the gray level run-length matrix (GLRLM), 5 texture features from the neighborhood gray tone difference matrix (NGTDM), and 744 texture features from wavelet transforms. Whereafter, seventy selected radiomics features, which were significantly different based on the independent-samples t-test or Mann-Whitney U test (P<0.05), were used in the LASSO regression. Binomial deviances and coefficients with different tuning parameters (λ) were shown in Figure 2A,2B, respectively. At the best λ value of 0.0165, LASSO analysis achieved optimal performance and yielded 14 candidate radiomic features. To determine the most effective subset and avoid overfitting, we constructed models using the top 5, 10, and all 14 features ranked by their absolute coefficient values, respectively, and found the model with the top 5 features demonstrated the best overall performance (Figure 2C). Thereafter, we used the LASSO model coefficients to calculate each patient’s Rad-score. A radiomics-clinical combined nomogram was then constructed on the basis of Rad-score, eosinophil percentage and CA19-9 (Figure 2D). ROC analysis showed that the radiomics-clinical combined model demonstrated superior diagnostic performance both in all three sets, outperforming the clinical and the radiomics model. Consistent findings were also observed in the PRC curves (Figure 2E-2J). Additionally, the combined model exhibited higher accuracy, sensitivity, and specificity, along with a superior F1 score and Youden index, indicating its enhanced predictive performance and discriminative ability. The thresholds were as follows: 0.493 for the clinical Model, 0.426 for the radiomics Model, and 0.421 for the combined Model. These values were applied directly to the internal and temporal validation cohorts for performance assessment. Table 4 summarizes the diagnostic efficacy of the clinical model, radiomics features, and the combined model in identifying fAIP versus PDAC in the training, validation, and temporal validation sets. In the temporal validation cohort, the combined model demonstrated sustained superiority, achieving an AUC of 0.88—surpassing both the clinical (AUC 0.80) and radiomics (AUC 0.84) models. It further exhibited optimal diagnostic performance with specificity (0.91) and the highest Youden index (0.70), confirming enhanced predictive reliability.

Figure 2 Model construction and performance evaluation in discriminating fAIP from PDAC. (A) Selection of the tuning parameter (λ) in the LASSO model via 5-fold cross validation based on the minimum MSE, plotted against log (λ). (B) LASSO coefficient path showing how feature coefficients shrink toward zero with increasing regularization strength. (C) Bar plot of the selected radiomic features and their absolute coefficient values, indicating feature importance. (D) Nomogram of the clinical-radiomics model for individualized prediction of PDAC probability. (E) ROC curves of the clinical, radiomics, and clinical-radiomics combined models in the training set. (F) ROC curves of the three models in the validation set. (G) ROC curves of the three models in the temporal validation set. (H) PRCs of the three models in the training set. (I) PRC of the three models in the validation set. (J) PRC of the three models in the temporal validation set. (K) DCA of the three models in the training set. (L) DCA of the three models in the validation set. (M) DCA of the three models in the temporal validation set. AUC, area under the curve; CA19-9, carbohydrate antigen 19-9; CI, confidence interval; CV, cross-validation; DCA, decision curve analysis; fAIP, focal autoimmune pancreatitis; GLCM, gray level co-occurrence matrix; GLSZM, gray level size zone matrix; HLH, high-low-high; LASSO, least absolute shrinkage and selection operator; LLH, low-low-high; LLL, low-low-low; MCC, maximal correlation coefficient; MSE, mean squared error; PDAC, pancreatic ductal adenocarcinoma; PRC, precision-recall curve; ROC, receiver operating characteristic.

DCA on the training set demonstrated that the combined model achieved higher net benefit than both the clinical and radiomics models across most threshold probabilities, with the advantage especially pronounced in the low-to-intermediate risk range (threshold <0.6). In the validation and temporal validation sets, the combined model’s net benefit remained superior to the clinical model and was comparable to that of the radiomics model (Figure 2K-2M). These results underscored the combined model’s key role in disease discrimination.

Model construction and performance evaluation in discriminating MFCP from PDAC

We then analyzed the serological profiles of the MFCP and PDAC groups (Table S3). While univariate analysis showed that higher levels of triglycerides and hemoglobin were associated with MFCP, and elevated conjugated bilirubin and CA19-9 levels were linked to PDAC (Table 5), multivariate analysis confirmed only triglycerides, conjugated bilirubin and CA19-9 as independent predictors.

Radiomics features significantly different between groups (P<0.05) based on t-tests or Mann-Whitney U tests were selected for LASSO regression. A total of 419 features were included, and 15 candidate features were retained at the optimal λ value of 0.0281 (Figure 3A,3B). To reduce overfitting, the five features with the highest coefficients were selected to construct the Rad-score for each patient. A combined model was constructed by integrating the selected radiomics features with key clinical variables. The predetermined classification thresholds were as follows: 0.306 for the clinical model, 0.441 for the radiomics model, and 0.448 for the combined Model. The diagnostic performance of the clinical, radiomics, and combined models was evaluated in the MFCP and PDAC training and validation sets (Figure 3C-3F). In the training set, the combined model achieved the highest AUC (0.96), outperforming both the clinical (0.87) and radiomics (0.92) models. In the validation set, the combined model remained superior with an AUC of 0.89, matching or exceeding the clinical (0.86) and radiomics (0.80) models. It also had the highest specificity (0.93) and Youden index (0.66), confirming its enhanced and reliable predictive performance (Table 6).

Figure 3 Model construction and performance evaluation in discriminating MFCP from PDAC. (A) MSE plotted against log₁₀(λ) during LASSO cross-validation. The optimal lambda value minimizes the MSE. (B) Bar plot of the selected radiomic features and their LASSO coefficients, indicating their relative contributions to classification. (C) ROC curves of the clinical, radiomics, and combined models in the training set. (D) ROC curves of the three models in the validation set. (E) PRC of the three models in the training set. (F) PRC of the three models in the validation set. (G) DCA of the three models in the training set. (H) DCA of the three models in the validation set. AUC, area under the curve; CV, cross-validation; DCA, decision curve analysis; GLSZM, gray level size zone matrix; HLH, high-low-high; LASSO, least absolute shrinkage and selection operator; LLH, low-low-high; LLL, low-low-low; MFCP, mass-forming chronic pancreatitis; MSE, mean squared error; PDAC, pancreatic ductal adenocarcinoma; PRC, precision-recall curve; ROC, receiver operating characteristic.

For the discrimination between MFCP and PDAC, the combined model demonstrated the highest net clinical benefit across the training and validation sets. Specifically, within a threshold probability range of approximately 0.20 to 0.60, its net benefit curve consistently surpassed those of the standalone clinical and radiomics models (Figure 3G,3H). This indicates that, within this clinical decision-making interval, using the combined model to guide management strategies yields a higher expected utility.

Model construction and performance evaluation in discriminating fAIP from MFCP

Using a similar approach, we next compared clinical characteristics and imaging data between fAIP and MFCP patients. Although the two groups showed some differences in laboratory results, multivariate analysis indicated that there were no significant differences in either serological marker (Tables S4,S5). Radiomic analysis alone demonstrated that LASSO analysis achieved optimal performance and yielded 7 candidate radiomic features at the best λ value of 0.01 (Figure 4A,4B), Rad-scores were calculated for these features and radiomic model was subsequent constructed. The predetermined classification threshold is 0.50 for the model. In the training cohort, the model showed good discriminative ability, with accuracy of 0.76, sensitivity of 0.77, AUC of ROC curve analysis of 0.79, and AUC of PRC analysis of 0.83 (Table 7), however, its performance declined markedly in the validation cohort (Figure 4C,4D).

Figure 4 Model construction and performance evaluation in discriminating fAIP from MFCP. (A) MSE plotted against log₁₀(λ) during LASSO cross-validation. The red dashed line marks the optimal lambda (λ=0.0100). (B) Bar plot of the selected radiomic features and their LASSO coefficients, indicating their relative contributions to classification. (C) ROC of the radiomics model in the training and validation sets. (D) PRC of the radiomics model in the training and validation sets. AUC, area under the curve; CV, cross-validation; fAIP, focal autoimmune pancreatitis; GLSZM, gray level size zone matrix; HHL, high-high-low; HLH, high-low-high; HLL, high-low-low; LASSO, least absolute shrinkage and selection operator; LLH, low-low-high; LHL, low-high-low; MFCP, mass-forming chronic pancreatitis; MSE, mean squared error; PRC, precision-recall curve; ROC, receiver operating characteristic.

Discussion

In this study, we developed a CT-based radiomics nomogram that integrates quantitative imaging features with key clinical parameters to differentiate fAIP from PDAC and MFCP in outpatients presenting with FPSLs. In the fAIP-versus-PDAC cohort, the combined clinical-radiomics model outperformed the standalone radiomics and clinical models, achieving an AUC of 0.95, 0.91 and 0.88 in training, internal validation, and temporal validation cohorts, respectively. Similar results were observed in the MFCP-versus-PDAC cohort. DCA demonstrated that the combined model provides a greater net clinical benefit than either the radiomics or clinical model alone—particularly across low to intermediate risk thresholds. Specifically, it showed a positive net benefit across a threshold probability range of approximately 0.20–0.60, which aligns well with real-world clinical decision points, when escalation to confirmatory testing (e.g., EUS-FNA) is typically considered. The findings suggest that the combined model demonstrates a favorable risk-benefit profile and may offer support for risk-stratified clinical decision-making in distinguishing FPSLs. The radiomics model showed initial promise in differentiating fAIP from MFCP in the training set (AUC =0.79), whereas its performance declined in the validation set (AUC =0.64). Taken together, our data suggested that integrated models offer superior utility compared to models based solely on imaging or clinical data in FPSL outpatients.

Increasing numbers of incidental FPSLs are being detected each year (1). PDAC is one of the most lethal forms of malignancy in FPSLs outpatients, with a global incidence and mortality rate of 8.1 and 6.9 per 100,000 people per year, respectively (12). For this reason, a widely adopted principle among clinicians is to approach all FPSLs with a presumption of malignancy until proven otherwise. However, up to 5.4–11% of PDAC cases are isoattenuating on traditional CT, making small, early-stage lesions likely to escape detection (13,14). This diagnostic imperative emphasizes the critical need for accurate characterization of a lesion’s malignant potential at the time of initial detection. Radiomics, also known as “imaging genomics”, offers promise in this area by extracting high-dimensional quantitative features from routine imaging, for example, texture and shape. These features can reveal tumor heterogeneity and subtle patterns that are not discernible to the human eye.

In this study, wavelet transform-based features accounted for a substantial proportion of the radiomics features selected by LASSO regression for modeling. Wavelet transform, as a multiscale and multidirectional texture feature extraction method, can effectively capture subtle yet consistent density variations in images, thereby enhancing the model’s ability to characterize lesion structural heterogeneity. In PDAC, such heterogeneity often corresponds to its characteristic tumor microenvironment (e.g., highly dense fibrous stroma) (15,16). These pathological alterations manifest as specific patterns of pixel intensity variation and spatial distribution on CT images, which can be quantitatively captured by wavelet transformation.

In the evaluation of outpatients with FPSLs, a central priority is to minimize missed diagnoses of PDAC. Triage tools in this context generally require a sensitivity of at least 0.80 to be considered clinically acceptable. The standalone clinical model, based on serum biomarkers such as CA19-9, exhibited high specificity but comparatively lower sensitivity in distinguishing PDAC from benign conditions. While this characteristic helps reduce false positives and avoid unnecessary invasive procedures, it also suggests a potential risk of missing early-stage, atypical, or minimally biomarker-elevated PDAC cases. These observations point to a clinical need for auxiliary tools that can enhance detection rates without substantially compromising specificity. To address this, we developed an integrated radiomics-clinical model. By combining imaging-derived heterogeneity features with conventional biomarkers, this model supports improved sensitivity while preserving diagnostic confidence.

Here, we observed high AUCs (~0.88–0.90) for distinguishing fAIP from PDAC using radiomics and enhanced model performance with incorporating clinical data, which was consistent with those reported in previous studies (7,17-19). One minor discrepancy lies in the AUC values: Lu’s model yielded AUCs of 0.83–0.87, whereas Park reported AUC of 0.975. This difference likely reflects variations in imaging protocols and patient selection. For example, Lu’s study focused on fAIP cases where the mass involved less than half of the pancreas, while Park’s cohort consisted strictly of pancreatic tumors, with AIP (including focal, multifocal, and diffuse forms) and PDAC cases closely matched by age and gender. Such controlled matching may have optimized feature extraction, whereas real-world heterogeneity likely contributes to lower AUCs in other studies. Importantly, none of the referenced studies found radiomics to underperform compared to human readers, while quantitative models were often noted to aid in differentiating diagnostically challenging cases. Overall, these comparisons support the plausibility of our results and reinforce the added value of clinical-radiomics integration.

Our radiomics model for MFCP versus PDAC demonstrated similarly strong performance in the training cohort (AUC =0.92), though validation performance was somewhat lower (AUC =0.80). However, the fusion model that integrated clinical data showed improved diagnostic accuracy, with AUCs of 0.96 in the training set and 0.89 in the validation set. The consistent success of integrated models across studies, such as an AUC of 0.825 from Qu and of 0.91 from Zhang (20,21), highlights their potential utility in resolving challenging differential diagnoses within FPSLs.

In contrast, direct radiomics comparisons between fAIP and MFCP remain scarce in the literature. Most existing studies focus on distinguishing each condition from PDAC, rather than from one another. This gap likely reflects clinical practice, where both fAIP and MFCP are benign inflammatory masses that often present with overlapping symptoms and fibrotic patterns in CT images. Our findings were unable to identify additional biochemical markers apart from IgG4, that could support the development of a reliable clinical discrimination model. This indicates the inherent limitations of conventional serological profiling in differentiating between these two inflammatory conditions. While elevated IgG4 levels are recognized as a specific diagnostic marker for Type 1 AIP, they are not always definitive. Imaging features such as diffuse, sausage-shaped pancreatic enlargement with delayed enhancement in AIP, or the duct penetration sign in MFCP can be helpful, but are not universally conclusive. Crucially, misdiagnosing AIP can lead to inevitable declines in quality of life and missed opportunities for surveillance of potential neoplastic transformation. This dual-risk scenario highlights the importance of accurate diagnosis before initiating corticosteroid treatment.

Our investigation represents the first systematic application of radiomic pipelines previously optimized for PDAC differentiation to this diagnostic challenge. Initial analysis within the training cohort demonstrated promising discriminatory performance (AUC =0.79). However, validation results were less favorable (AUC =0.64). This can be attributed to the fact that both fAIP and MFCP are benign entities primarily characterized by chronic inflammation and dense fibrosis, resulting in relatively homogeneous enhancement on venous-phase CECT and insufficient textural contrast for radiomics feature extraction. Consequently, radiomics based on single-phase venous CECT appears more suitable for distinguishing benign from malignant lesions, while its capability to differentiate between benign fibroinflammatory diseases remains limited. Given the lack of comparable studies, and even expert radiologists may struggle to distinguish them without serological data or a therapeutic trial of steroids, our modest performance in this task is both expected and reasonable. Future studies may benefit from incorporating multi-phase CT, MRI texture analysis, or functional imaging to further improve the differentiation of benign pancreatic diseases.

The primary strength of this study lies in the development and internal validation of a clinical–radiomics fusion nomogram for the early diagnosis of FPSLs. Despite encouraging results, several limitations warrant consideration. First, this retrospective, single-center study lacked multi-center external validation, introducing potential selection bias and limiting generalizability, particularly for the fAIP-MFCP task. Second, the relatively small sample size of the fAIP and MFCP subgroups may limit the events-per-variable (EPV) ratio and contribute to potential model instability, especially in high-dimensional feature selection processes such as LASSO. Third, we included only patients with definitive diagnoses, leaving the model’s performance on more ambiguous or borderline cases untested. Finally, lesion segmentation was performed manually, which may introduce inter-operator variability. Given these limitations, this work should be considered an exploratory proof-of-concept that requires further validation.

Conclusions

In summary, this study preliminarily developed and internally validated a clinical-radiomics fusion model that shows potential to aid in the early diagnosis of FPSLs, particularly in differentiating benign inflammatory masses from PDAC in an outpatient setting. Our results support that the continued integration of AI-driven image biomarkers with clinical variables may offer a complementary approach to current diagnostic pathways for FPSLs diagnosis. Future research should focus on external validation, incorporation of multi-modal data, and prospective studies to fully realize the clinical impact of these advanced diagnostic approaches in improving outcomes for outpatients with FPSLs.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-2017/rc

Data Sharing Statement: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-2017/dss

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-2017/prf

Funding: The study was supported by Hubei Province Health and Family Planning Scientific Research Project (No. WJ2023M014 to B.W.), Open Project of the Key Laboratory of Organ Transplantation, Ministry of Education and NHC (No. 2020QYKF02 to B.W.), and Tongji Hospital Scientific Research Fund Project (No. 2023A14 to B.W.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-2017/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committee of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology (approval No. TJ-IRB202409044) and individual consent for this analysis was waived due to the retrospective nature.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Best LM, Rawji V, Pereira SP, et al. Imaging modalities for characterising focal pancreatic lesions. Cochrane Database Syst Rev 2017;4:CD010213. [Crossref] [PubMed]
2. Tornel-Avelar AI, Velarde Ruiz-Velasco JA, Pelaez-Luna M. Pancreatic cancer, autoimmune or chronic pancreatitis, beyond tissue diagnosis: Collateral imaging and clinical characteristics may differentiate them. World J Gastrointest Oncol 2023;15:925-42. [Crossref] [PubMed]
3. Elsherif SB, Virarkar M, Javadi S, et al. Pancreatitis and PDAC: association and differentiation. Abdom Radiol (NY) 2020;45:1324-37. [Crossref] [PubMed]
4. Mastrantoni L, Chiaravalli M, Spring A, et al. Comparison of first-line chemotherapy regimens in unresectable locally advanced or metastatic pancreatic cancer: a systematic review and Bayesian network meta-analysis. Lancet Oncol 2024;25:1655-65. [Crossref] [PubMed]
5. Löhr JM, Beuers U, Vujasinovic M, et al. European Guideline on IgG4-related digestive disease - UEG and SGF evidence-based recommendations. United European Gastroenterol J 2020;8:637-66. [Crossref] [PubMed]
6. Harne PS, Harne V, Wray C, et al. Endoscopic innovations in diagnosis and management of pancreatic cancer: a narrative review and future directions. Therap Adv Gastroenterol 2024;17:17562848241297434. [Crossref] [PubMed]
7. Park S, Chu LC, Hruban RH, et al. Differentiating autoimmune pancreatitis from pancreatic ductal adenocarcinoma with CT radiomics features. Diagn Interv Imaging 2020;101:555-64. [Crossref] [PubMed]
8. Yu H, Huang Z, Li M, et al. Differential Diagnosis of Nonhypervascular Pancreatic Neuroendocrine Neoplasms From Pancreatic Ductal Adenocarcinomas, Based on Computed Tomography Radiological Features and Texture Analysis. Acad Radiol 2020;27:332-41. [Crossref] [PubMed]
9. Deng Y, Ming B, Zhou T, et al. Radiomics Model Based on MR Images to Discriminate Pancreatic Ductal Adenocarcinoma and Mass-Forming Chronic Pancreatitis Lesions. Front Oncol 2021;11:620981. [Crossref] [PubMed]
10. Bush N, Khashab M, Akshintala VS. Current and Emerging Applications of Artificial Intelligence (AI) in the Management of Pancreatobiliary (PB) disorders. Curr Gastroenterol Rep 2024;26:304-9. [Crossref] [PubMed]
11. Podină N, Gheorghe EC, Constantin A, et al. Artificial Intelligence in Pancreatic Imaging: A Systematic Review. United European Gastroenterol J 2025;13:55-77. [Crossref] [PubMed]
12. Xiao AY, Tan ML, Wu LM, et al. Global incidence and mortality of pancreatic diseases: a systematic review, meta-analysis, and meta-regression of population-based cohort studies. Lancet Gastroenterol Hepatol 2016;1:45-55. [Crossref] [PubMed]
13. Lim CY, Min JH, Hwang JA, et al. Assessment of main pancreatic duct cutoff with dilatation, but without visible pancreatic focal lesion on MDCT: a novel diagnostic approach for malignant stricture using a CT-based nomogram. Eur Radiol 2022;32:8285-95. [Crossref] [PubMed]
14. Abunahel BM, Pontre B, Kumar H, et al. Pancreas image mining: a systematic review of radiomics. Eur Radiol 2021;31:3447-67. [Crossref] [PubMed]
15. Yun G, Kim YH, Lee YJ, et al. Tumor heterogeneity of pancreas head cancer assessed by CT texture analysis: association with survival outcomes after curative resection. Sci Rep 2018;8:7226. [Crossref] [PubMed]
16. Li X, Ma Q, Xu Q, et al. Targeting the cancer-stroma interaction: a potential approach for pancreatic cancer treatment. Curr Pharm Des 2012;18:2404-15. [Crossref] [PubMed]
17. Ziegelmayer S, Kaissis G, Harder F, et al. Deep Convolutional Neural Network-Assisted Feature Extraction for Diagnostic Discrimination and Feature Visualization in Pancreatic Ductal Adenocarcinoma (PDAC) versus Autoimmune Pancreatitis (AIP). J Clin Med 2020;9:4013. [Crossref] [PubMed]
18. E L. Differentiation of Focal-Type Autoimmune Pancreatitis From Pancreatic Ductal Adenocarcinoma Using Radiomics Based on Multiphasic Computed Tomography. J Comput Assist Tomogr 2020;44:511-8. [Crossref] [PubMed]
19. Lu J, Jiang N, Zhang Y, et al. A CT based radiomics nomogram for differentiation between focal-type autoimmune pancreatitis and pancreatic ductal adenocarcinoma. Front Oncol 2023;13:979437. [Crossref] [PubMed]
20. Qu W, Zhou Z, Yuan G, et al. Is the radiomics-clinical combined model helpful in distinguishing between pancreatic cancer and mass-forming pancreatitis? Eur J Radiol 2023;164:110857. [Crossref] [PubMed]
21. Zhang H, Meng Y, Li Q, et al. Two nomograms for differentiating mass-forming chronic pancreatitis from pancreatic ductal adenocarcinoma in patients with chronic pancreatitis. Eur Radiol 2022;32:6336-47. [Crossref] [PubMed]