Pancreatic solid pseudopapillary neoplasm (PSPN) is a rare, low-grade malignant tumor accounting for 0.9% – 2.7% of all pancreatic tumors. Previous studies have shown that PSPN is more common in women under 40 years of age, whereas in men, the incidence of PSPN is lower, the age of onset is much older, and the malignant grade is higher (1–5). PSPN is a heterogeneous tumor, with a minor subset potentially exhibiting invasive characteristics, which significantly influence patient prognosis. Invasive behavior may involve perineural invasion, infiltration of adjacent organs and blood vessels, invasion into the pancreas and surrounding adipose tissue, as well as distant metastasis and regional lymph node involvement (6, 7). Currently, surgical resection is considered the preferred and most effective treatment for PSPN, and radical surgery is associated with a favorable prognosis, achieving a postoperative survival rate of 80% – 90% (8–10). In recent years, conservative surgical methods, such as tumor exenteration or laparoscopic surgery, have often been used for noninvasive PSPN patients, whereas for invasive PSPN patients, extended radical resection is needed to ensure a satisfactory long-term prognosis, and incomplete resection or positive resection margins may increase the risk of recurrence (11–13). Compared with conservative surgical methods, a wide range of pancreatic resections may lead to postoperative complications and pancreatic endocrine and exocrine dysfunction. Consequently, the preoperative differentiation of invasive and noninvasive PSPN is highly important for selecting appropriate surgical methods in the clinic and avoiding unnecessary resection (14, 15). However, obtaining preoperative pathological results is often challenging. The accuracy of pathological diagnosis from needle biopsy is restricted by the sample quality and quantity, which may not adequately reflect tumor heterogeneity. Additionally, needle biopsy carries the risk of tumor cell dissemination along the needle tract. Consequently, this complicates and raises controversy regarding the selection of the appropriate surgical approach for surgeons (16, 17).

CT is the first choice for the examination of pancreatic lesions, and enhanced CT can display invasive signs well, which is highly valuable for the preoperative evaluation of invasive and noninvasive PSPN (2). Nevertheless, image interpretation is limited by the subjectivity of radiologists; thus, finding a quantitative and objective method to evaluate invasive and noninvasive PSPN is highly important. Radiomics uses high-throughput extraction of implicit quantitative features (texture, shape, wavelet transform parameters, etc.) in images, combined with machine learning algorithms to mine tumor heterogeneity information, providing a new idea for non-invasive prediction of PSPN invasiveness. Previous studies have shown that radiomics has great potential in the diagnosis, assessment of invasiveness, biochemical recurrence prediction, and metastasis of pancreatic tumors (18–21). Tobaly et al. (18) found that radiomics features extracted from CT images can distinguish low-risk and high-risk Intraductal papillary mucinous neoplasms and guide surgical decision-making. Shi et al. (20) extracted radiomics features based on MR images and constructed a model combining clinical information to differentiate PSPN and pancreatic neuroendocrine tumor. The study results showed that the accuracy of the radiomics model was 92.42%, which was significantly higher than that of subjective diagnosis.

However, previous studies have focused on differentiating PSPN from other pancreatic tumors. Few studies have employed contrast-enhanced CT (CECT) radiomics to predict the invasiveness of PSPN. The purpose of this study was to developed and demonstrated a radiomics nomogram based on radiomics and traditional CT signs to predict the invasiveness of PSPN and provide clinicians with treatment decisions, especially the choice of surgical approach.

In this study, predictive models incorporating radiomics and traditional CT indicators were developed to differentiate between invasive and non-invasive PSPN. Among all the models, the radiomics nomogram integrating traditional CT signs with radiomics signature demonstrated the highest diagnostic performance, with AUCs of 0.874 and 0.867 in the training and validation sets, respectively. This performance surpassed that of both the traditional CT model and the radiomics models. These findings suggest that the combined use of radiomics and traditional CT signs enhances classification accuracy, potentially aiding in clinical treatment decision-making.

Previous studies (22, 23) have shown that sex, age, tumor shape calcification, pancreatic atrophy, and pancreatic duct dilation are not significantly different between invasive and noninvasive PSPN patients, and similar results were obtained in our study. Liang et al. (24) indicated that the location of the tumor varied significantly between invasive and noninvasive PSPN, with invasive PSPN more frequently found in the pancreatic tail. However, the present study showed no statistically significant difference in tumor location. The discrepancy may be attributed to the different grouping methods used; in this study, the pancreas was divided into two groups, with the head and neck classified together and the body and tail grouped separately, which differs from the classification used in previous research. Wang et al. (8) reported that tumor size is an independent predictor of invasiveness in PSPN patients. Kim et al. (25) reported that when the tumor diameter was greater than 5 cm, the probability of presenting with invasive PSPN significantly increased (P = 0.022). However, to date, there is no consensus on the ability of tumor size to predict invasive PSPN. In our study, tumor size was statistically significant only in the validation set but not in the total sample, and there may be bias due to the small sample size in the validation set. Further large sample studies are needed for verification in the future. Wang et al. (8) and Rastogi et al. (26) reported no significant differences in the solid component between invasive and noninvasive PSPN, whereas in our study, invasive PSPN predominantly presented as solid or mixed solid-cystic tumors, and the solid component was significantly different (P<0.05). This finding suggests that tumor invasive behavior may be related to the extent of solid components, possibly because increased vascularity within the tumor leads to invasive growth. Moreover, the difference in tumor margins between invasive PSPN and noninvasive PSPN was statistically significant, possibly because PSPN has a richer blood supply and is more likely to have external growth, resulting in focal discontinuity of the capsule and invasion of the pancreas and surrounding adipose tissue, which results in unclear tumor boundaries.

Radiomics is an emerging discipline of high-throughput extraction of quantitative features from medical images and the use of these features to build models for clinical decision-making with the goal of improving diagnostic accuracy (27–29). In recent years, radiomics has shown good advantages in the field of medicine (30–32). However, there has been limited research utilizing multiphase CT radiomics to predict the invasiveness of PSPN; only Huang et al. (33) used dual-phase contrast-enhanced imaging to predict the invasiveness of PSPN. However, fewer cases were included in Huang’s study than in our study, and the study lacked further verification and did not explore the significance of unenhanced phases and traditional CT signs. In this study, the combination of radiomics and traditional CT signs to construct a radiomics nomogram may fill this gap.

In our study, 14 valuable radiomics features were extracted from the four radiomics models, most of which were first-order and texture features from wavelet-transformed images. The optimal features of the four models all included Wavelet_LLL_NGTDM_Busyness features. The high Busyness values in NGTDM features reflects the heterogeneity between the lesion pixel local grayscale and adjacent pixel heterogeneity. Invasive PSPN has a higher Busyness value than non-invasive PSPN. The invasive PSPN may have a more complex texture due to necrosis, neovascularization or uneven cell density, which is manifested as a higher Busyness value. The Dependence Variance of GLDM is an important feature for quantifying the heterogeneity of local gray-scale dependency in images. It is jointly influenced by cell density, interstitial proportion and vascular distribution. The variation of Dependence Variance value may be used to quantify the complex microenvironment within invasive PSPN. The size zone non-uniformity values in GLSZM features is positively correlated with the invasiveness of PSPN. High size zone non-uniformity value indicates represent a non-uniform texture, that is, high heterogeneity, and can be used as a powerful indicator to predict the invasiveness of PSPN. In addition, the skewness and kurtosis values extracted from wavelet transform images have significant weights and higher values in invasive PSPN, reflecting the complexity and heterogeneity of the cell density distribution in invasive PSPN, possibly due to rich abnormal vascular formation, changes in cell permeability, and necrosis, resulting in mixed internal tumor components and complex grayscale distributions. These findings suggest that the internal structure and heterogeneity of tumors are closely related to invasiveness. The above results indicate that radiomics features are potential noninvasive biomarkers for predicting the aggressiveness of PSPN.

Wang et al. (8) reported that CT imaging findings may help distinguish invasive PSPN from noninvasive PSPN, and the AUC was 0.77. In the current work, the diagnostic efficacy of the radiomics model was better than that of Wang’s study, indicating that radiomics contributes greatly to diagnosis. In a study by Rastogi et al. (26) in which CECT was used to predict PSPN invasiveness, invasive PSPN showed a greater degree of enhancement than noninvasive PSPN did in the delayed phase, whereas the arterial and venous phases were not significantly different. In this study, all three single-phase models were capable of predicting PSPN invasiveness, with the arterial-phase model exhibiting the highest performance. This may be attributed to the fact that the increased blood supply in invasive PSPNs is more effectively captured in arterial-phase images. Furthermore, invasive PSPNs are more prone to vascular invasion, resulting in compensatory increases in peripheral blood supply, which enhances the visibility in arterial-phase imaging. Additionally, due to the presence of mixed substances such as hyaluronic acid and collagen in the intercellular stroma, enhancement of the interstitial component may be observed during the venous phase due to contrast agent inflow, while vascular enhancement may be less pronounced. In contrast, arterial-phase images primarily show enhancement of tumor blood vessels, without significant enhancement of the interstitial component (33). These findings may indicate the predictive ability of radiomics features on the basis of the arterial phase in predicting tumor invasiveness. We subsequently combined the three phases to obtain a combined model, which significantly improved the predictive ability for PSPN invasiveness compared with single-phase models.

Radiomics is not the only determinant of diagnosis, and when radiomics is combined with other relevant data, more reliable and accurate results can be produced (34). Song et al. (35) constructed a nomogram combining age and MRI radiomics features, which was used to differentiate hypovascular nonfunctional pancreatic neuroendocrine tumors from PSPN. The AUC of the nomogram in the validation set was 0.920, which was greater than that of the radiomics model (AUC = 0.907). Gu et al. (36) demonstrated that combining clinical information with radiomics improved the differentiation of PSPN from three other conditions, namely adenocarcinoma, neuroendocrine tumors, and pancreatic cystadenomas (AUC = 0.962). Liang et al. (24) utilized radiomics features derived from T1-weighted imaging, T2-weighted imaging, diffusion-weighted imaging, and contrast-enhanced T1WI sequences, integrating them with clinical data to develop a radiomics nomogram for distinguishing between invasive and noninvasive PSPN, and the radiomics nomogram showed the best diagnostic performance (AUC = 0.808). In the present study, traditional CT indicators were combined with radiomics features to establish a radiomics nomogram, achieving an AUC of 0.867 in the validation set, surpassing both the traditional CT and radiomics models. When compared with Liang et al. (24), the proposed model showed enhanced predictive performance, suggesting that CECT may be more effective than MRI in assessing PSPN invasiveness. To validate the clinical applicability of the nomogram model, DCA was performed, indicating that within the specified threshold probability range, this model provided greater net benefit for clinical decision-making.

The primary objective of radiomics is to facilitate classification and prediction. Diagnostic and classification models in radiomics research predominantly rely on machine learning (ML) techniques, such as logistic regression analysis, random forests, support vector machines, LASSO, and linear discriminant analysis, with logistic regression being the most widely used. In this study, only the logistic regression model was employed to assess PSPN invasiveness, and the efficacy of other algorithms in similar ML tasks remains uncertain and warrants further exploration.

This study has several limitations. First, the retrospective nature of the analysis, conducted across multiple centers, may introduce selection bias in patient cohorts. Second, manual delineation of ROIs by different readers could lead to bias, potentially impacting the reliability of radiomics features. Although features with ICC > 0.75 were selected to mitigate this issue, there is a critical need for automated and accurate tumor segmentation techniques. Thirdly, This study only used 2D ROI analysis, which may ignore the spatial heterogeneity of lesions, while 3D ROI retains the three-dimensional spatial information of lesions and can extract more abundant radiomics features. In the future, we will further expand the data and discuss the advantages and disadvantages of 2D ROI and 3D ROI. Finally, due to the lack of external validation, we cannot ensure that our model has the same diagnostic performance when dealing with external datasets.

In conclusion, Our study shows that combining traditional CT signs with radiomics and constructing a radiomics nomogram based on multi-phase contrast-enhanced CT can effectively predict the invasiveness of PSPN, which may aid in clinical decision-making for treatment strategies.