Pancreatic ductal adenocarcinoma (PDAC) accounts for the seventh-highest cancer mortality rate worldwide, and 50% of patients with PDAC suffer from metastatic disease (1). Recently, neoadjuvant therapy has emerged as a significant component in the management of PDAC (2). According to an investigation, the administration of neoadjuvant therapy followed by resection is associated with improved rates of overall survival in comparison to individuals who undergo upfront resection followed by adjuvant therapy (3). Neoadjuvant chemotherapy for pancreatic cancer with significant lymph node (LN) metastases was recommended by the National Comprehensive Cancer Network (NCCN) guidelines (4). Hence, the preoperative evaluation of LN status holds significant importance in establishing a personalized treatment strategy for individuals diagnosed with PDAC.

The most commonly used imaging modality to evaluate patients with PDAC prior to surgery is computed tomography (CT). However, previous studies have suggested that the ability of CT to effectively forecast the presence of nodal involvement in surgically treatable PDAC is constrained (5). Approximately 70% of patients with negative clinical lymph nodes present with lymph node involvement upon pathologic assessment (6). The lymph node enlargement may occur independently of localized inflammation caused by malignant biliary obstruction (7). This indicates that the preoperative radiological staging of nodule involvement remains difficult. Therefore, there is an unquestionable need for more effective methods to assess the LN status for patients with pancreatic cancer.

Radiomics provides a large amount of medical image information that can reveal hidden features of diseases that are invisible to the naked eye (8). Previous research has demonstrated the efficacy of integrating radiomics and machine learning (ML) techniques in many applications (9–11). Several researchers have constructed multiphasic contrast-enhanced CT (CECT) radiomics models to evaluate the preoperative LN status of PDAC (12–14). These studies indicated that radiomics models have significant potential in predicting pancreatic cancer with lymph node metastasis (LNM). However, the researchers predominantly chose modeling algorithms based on their preferences and limitations in knowledge. It is imperative that evidence be provided to select appropriate models for solving clinical problems. To the best of our knowledge, no study has explored the use of radiomics combined with multiple ML algorithms in the preoperative identification of the LN status in PDAC. The objective of this study was to evaluate the predictive value of clinical data and radiomic features for forecasting the LN status of patients with pancreatic cancer using a combination of various ML algorithms.

PDAC has a heightened mortality rate and a dismal prognosis, mostly attributed to the challenges associated with early identification and the curative treatment modalities. In patients with PDAC, pancreatectomy has the greatest likelihood of improving long-term survival. LN status as an essential and independent prognostic indicator has been demonstrated (18). However, there remains a debate as to whether standard or extended LN dissection should be included during pancreatectomy (19, 20). Therefore, it is essential to accurately stage the LN before surgery in order to provide patients with comprehensive counseling on surgical choices and prognosis. The objective of this research was to construct and verify multiple ML models using CECT to differentiate the LN status in PDAC. As the combined model incorporated elements like CA199 levels and tumor size, it varied from the conventional single radiomics model. The combined model demonstrated significant capability of distinguishing LN metastasis in the training (AUC, 0.920) and validation (AUC, 0.866) cohorts. To the best of our knowledge, this is the first study to compare multiple ML algorithms to detect the LN metastases in PDAC preoperatively.

CA199 is currently regarded as the most appropriate serum marker for patients with resectable PDAC, as it is the only prognostic biomarker approved by the Food and Drug Administration (21). A study has provided evidence of the clinical significance of the preoperative CA199 level as a prognostic marker in patients with PDAC presenting with metastases in the para-aortic LN (22). In our study, the CA199 level was validated as an independent risk factor of LN metastasis for PDAC. We speculated that such results may be because CA199 is a particular biomarker that reflects the biological activity and that the greater the CA199 level, the worse the patient’s prognosis (23). A study found a positive association between the size of tumors and occurrence of LN metastases in endometrial carcinoma (24). Similarly, our investigation has unveiled that the size of the tumor serves as a noteworthy and independent predictor of the occurrence of LN metastasis. However, clinical models consisting of CA199 and tumor size have poor predictive ability for LN status (AUC= 0.714 in training datasets; AUC = 0.657 in validation datasets). Some studies suggested that patients with severe obstructive jaundice may have elevated CA199, which can mislead predictive results (25). It partially explains the poor predictive efficacy of clinical models. The predictive value of inflammatory indices in relation to the prognosis of individuals diagnosed with PDAC has been shown in previous studies (26, 27). However, our investigation could not confirm the association between inflammatory indices and LN metastases. The differences in the study population may explain the inconsistencies with previous findings, and research with larger sample sizes or in other populations is warranted in the future.

Although CT is the most commonly used method to evaluate pancreatic cancer resectability, Imai et al. (28) indicated that low-order data such as LN diameter and volume measured from CT image cannot reflect obvious differences between patients with pancreatic cancer with and without LN metastasis. This may explain why visual assessment of CT scans has a low efficiency for detecting LN metastases among patients with pancreatic cancer. Radiomics, also referred to as a “whole-tumor virtual biopsy technique,” enables the extraction of numerous image features from the entire tumor, reflecting its heterogeneity and characteristics (29). A biopsy examines only a portion of the tumor tissue, which cannot comprehensively assess the intra-tumor heterogeneity before surgery (30). In contrast, radiomics can non-invasively reflect comprehensive information about the entire tumor. This approach has promise for enhancing diagnostic capabilities and facilitating the development of personalized treatment approaches (31). The data from radiomics contain first-, second-, and higher-order features that can hardly be identified by the naked eye (31). This data was used for predicting the LN status of bladder, breast, and gastric, and esophageal cancers and were suggested to have some advantages (32–35). Several studies in recent years have used CECT radiomics analysis of primary lesions to predict LN metastasis. PDAC tumors or peri-tumors have been shown to have a high lymphatic vessel density (36, 37). PDAC commonly infiltrates the LNs through the lymphatic system, as opposed to direct or adjacent extension of the primary cancer into the LN (38, 39). Accordingly, analyzing the image of the primary tumor may be more effective than the image of LNs in predicting LN metastasis. That is why ROIs were segmented into tumor regions in this study. Some previous studies have used only one VP radiomics feature to construct the models; however, our study shows that AP radiomics feature is also important (12, 40, 41). Some other studies have only considered AP images, which may miss some useful information about the VP images (7, 42). Our findings suggest that AP and VP radiomics feature have comparable performances in predicting the LN status. Besides, the predictive performance of the AP+VP-Radscore constructed by combining AP and VP features was higher than that of the model constructed by a single phase. Thus, incorporating more CT phases may allow for the development of more effective models. Shi et al. conducted a study wherein they employed MRI radiomics to detect LN metastasis in individuals with PDAC prior to surgery (43). However, it should be noted that their study incorporated a limited number of clinical factors. A previous study developed a radiomics-based strategy for preoperative prediction of LN status in patients with resectable PDAC. However, only the largest tumor strata were examined in their investigation, and the sample comprised only 85 cases (41). Most of these previous studies used only one algorithm, hence limited conclusions can be drawn. Our study differed from previous studies as we incorporated AP and VP of CT rather than one scan phase. Multiple radiomics ML models were compared rather than solely focusing on a single model. The integration of the clinical variables and the AP+VP-Radscore in our combined model demonstrated a high level of predictive capability, as evidenced by an AUC of 0.866 in the validation cohort. This performance surpassed that of both the single phase Radscore and clinical model. Hence, the inclusion of clinical variables holds significance and warrants consideration in the analysis of radiomics.

In this study, the predictive power of the combined AP and VP features outperformed that of separate AP or VP model. The AP+VP-Radscore in our study was calculated by six wavelet texture, two LoG filtered texture, and one original features. Wavelet features have been confirmed as useful indexes in predicting the tumor pathological grade and histologic subtype (44, 45). According to our study, the wavelet texture features took an important position in the LN metastasis prediction model. By using the wavelet decomposition of images at several scales, fine and coarse textures could be extracted to depict the spatial heterogeneity within tumor regions (46). The findings of this research align with those of other studies that employed wavelet-based features in their radiomics models. A previous research endeavor was initiated to construct a predictive model to assess the preoperative LN status for intrahepatic cholangiocarcinoma, which was constructed using wavelet features exclusively (47). In another research aimed at evaluating the role of radiomics in predicting LN status in intrahepatic cholangiocarcinoma, it was observed that wavelet features constituted 79% of the features incorporated into the model (48). Xiao et al. conducted research to establish and validate a nomogram for the purpose of forecasting the occurrence of LN metastasis in individuals with early-stage cervical cancer. The constructed model encompassed a total of 23 features, with 19 of these being wavelet features (49). We speculated that the possible explanation for this phenomenon could be the potential correlation between wavelet features and the infiltration of malignancies into the lymphatic system. LoG filtering is a sophisticated technique for picture filtering that integrates the principles of Laplace and Gaussian filtering. The utilization of Laplacian filtering in image processing can effectively emphasize locations within the image that have undergone mutations in gray level, hence enhancing the contrast in gray levels. The application of a Gaussian operator can mitigate the noise introduced by the Laplace operator (50, 51). LoG filtering can also be beneficial in detecting LN metastases of PDAC based on the results of our study. Our results show that GLSZM features are helpful in diagnosing metastatic LNs in PDAC. A previous study used radiomics to predict LN metastasis in individuals diagnosed with PDAC, and the model also included GLSZM features (12). GLSZM features quantify gray level zones in an image. The texture features based on GLSZM, which capture the interplay of adjacent pixels, have demonstrated a superior ability to quantify tumor texture and heterogeneity compared to features based on histograms (52).

ML algorithms, including SVM, LDA, and LR, have been extensively employed in the domain of clinical prediction and classification tasks (53, 54). Multiple studies show that LR has better performance than other ML algorithms (55–58). A meta-analysis which compared the efficacy of multiple ML algorithms in pregnancy care prognosis prediction indicated that LR models had superior performance compared to non-LR models in trials with a low risk of bias (57). Li et al. (58) employed stepwise selection using LR as well as four ML techniques, including gradient boosting machine, RF, XGBoost, and single-layer neural networks, in order to construct predictive models for hip fractures. LR was found to be best in terms of discrimination performance. ML has demonstrated high performance in tasks characterized by a significant signal-to-noise ratio, such as identification of handwriting and video games. Clinical predictive issues frequently exhibit one low signal-to-noise ratio (59). LR is based on a sigmoid function and works best on binary classification problems (60). Consequently, LR with universality cannot be ignored in this study. To explore the best algorithm for predicting PDAC LN status, we compared 233 models and finally confirmed that the Lasso-Logistic regression model constructed with AP+VP features outperformed other ML algorithms. This discovery implies that the efficacy of the model is impacted by the characteristics of the algorithm and its congruence with the study goals.

Our study had several limitations. First, it was only based on single-center data. Although our study adopted cross-validation, it may not avoid the overfitting risk. A large sample, multicenter study utilizing multiple CT scanners is necessary to validate the accuracy and stability of our combined model. Second, subgroup analysis based on sites of LN metastases was not implemented. Third, we only evaluated the AP and VP images of CECT; combining the plain and delayed phases may contain more information about tumor heterogeneity. Moreover, this study did not employ dynamic radiomics features to quantify the degree of variation between AP and VP, thereby overlooking the temporal evolution of radiomics features. Dynamic radiomics features should also be explored in future studies. Fourth, we used the default parameter settings of ML algorithms instead of tuning the parameters, which may have hindered the algorithm from achieving its optimal performance. Finally, the image segmentation approach utilized in this study relied on manual delineation. In the future, automated segmentation techniques could be employed to enhance consistency and efficiency.

In summary, of the 233 radiomics models examined, the model built by applying AP+VP radiomics features and a combination of Lasso-Logistic algorithm had the most favorable performance in both the training and validation cohorts. Our investigation showed that integrating the AP+VP-Radscore with clinical parameters yielded the best performance. This combined model has the potential to serve as an accurate and non-invasive instrument for forecasting LN metastasis of PDAC, hence facilitating clinical decision-making.