Development and validation of machine learning models integrating spectral computed tomography-derived three-dimensional quantitative parameters for predicting histopathological grade in pancreatic ductal adenocarcinoma

Introduction

Pancreatic ductal adenocarcinoma (PDAC) accounts for 85–95% of pancreatic cancer cases and has an extremely poor prognosis, with 5-year survival rates remaining below 10% (1,2). Research has demonstrated that the pathological differentiation grade of PDAC is closely related to the choice of treatment modalities, therapeutic efficacy, and prognosis (3,4). For well-differentiated PDAC, radical resection can achieve relatively longer survival. Conversely, for poorly differentiated PDAC, postoperative survival does not significantly improve, and postoperative complications may lead to a decline in the quality of life for patients (5). Therefore, accurate preoperative grading of tumors may be the key to individualized and precise treatment for PDAC.

The degree of tumor differentiation in PDAC is often assessed only through the comprehensive evaluation of postoperative specimens in clinical practice, which limits its ability to guide preoperative clinical decision-making. Due to limited tissue sampling and tumor heterogeneity, endoscopic ultrasound-guided fine needle aspiration and percutaneous biopsy, as invasive tools, are insufficient for accurately assessing tumor differentiation (6,7). Several biomarkers have been explored but remain insufficient. For instance, while serum carbohydrate antigen 19-9 (CA19-9) is a widely used biomarker, its correlation with tumor grade is inconsistent (8). Conventional contrast-enhanced computed tomography (CT) remains the standard imaging modality for PDAC staging. Wang et al. (9) reported a negative correlation between CT enhancement patterns and pathological grade (r=−0.784) in a cohort of 54 patients with pancreatic cancer. However, this finding should be interpreted cautiously because of the relatively small sample size and potential bias. Furthermore, its assessment of tumor grade is largely subjective and qualitative, based on morphological features like necrosis, which fail to capture the underlying biological heterogeneity (10).

Spectral CT, also known as dual-energy CT, overcomes these limitations by enabling the quantification of material-specific parameters beyond conventional CT attenuation values, such as iodine concentration (IC) and the effective atomic number (Zeff) (11). These parameters provide indirect, noninvasive information about tumor vascularization, cellular density, and metabolic status. Emerging evidence indicates that quantitative parameters derived from energy-spectrum CT correlate with histopathological differentiation or cancer stage in PDAC (10,12). However, most prior studies have relied on measurements from a single two-dimensional region of interest (ROI) placed on the largest tumor slice. This approach is inherently limited, as it fails to account for the well-recognized three-dimensional (3D) heterogeneity of PDAC, which can lead to sampling bias and poor reproducibility. In contrast, 3D volumetric analysis of the entire tumor provides a more comprehensive and reproducible assessment of tumor heterogeneity (13,14). Preliminary work has shown the potential of 3D spectral CT parameters in predicting the Ki-67 proliferation index in PDAC (15), but their role in predicting histopathological grade, especially when combined with clinical data, remains largely unexplored.

Machine learning (ML) offers a powerful framework for integrating diverse datasets and uncovering complex, non-linear relationships that traditional statistical methods may miss. ML algorithms have been widely used to discern key risk factors and prognostic markers for various diseases (16-19). For this study, we deliberately selected three algorithms with distinct theoretical foundations to balance predictive performance and interpretability. Logistic regression (LR), a linear model chosen for its high interpretability, allowing direct assessment of the direction and magnitude of each feature’s association with tumor grade (20). Light Gradient Boosting Machine (LightGBM), a tree-based ensemble method selected for its efficiency in handling high-dimensional data and capturing complex feature interactions (21). Multilayer Perceptron (MLP), a neural network model chosen for its capacity to model highly intricate and non-linear relationships (22). This study therefore aimed to develop and evaluate multiple ML algorithms that integrate spectral CT-derived 3D quantitative parameters with clinical features for the preoperative prediction of histopathological grade in patients with PDAC. We present this article in accordance with the TRIPOD reporting checklist (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2025-1-1023/rc).

Methods

Patients selection

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Institutional Review Board of the Chongqing General Hospital (No. KY S2024-069-01), individual consent for this retrospective analysis was waived. This was a retrospective diagnostic model development and validation study conducted at a single tertiary referral center. Between August 2021 and February 2025, this study retrospectively enrolled 285 patients with pathologically confirmed PDAC and documented histological grades who underwent preoperative spectral CT for pretreatment assessment. Patients were excluded based on the following criteria: (I) patients who had received any anticancer therapy prior to the spectral CT examination (n=134); (II) patients with suboptimal image quality (n=11); (III) patients with missing or incomplete clinical data (n=8); (IV) patients with concurrent primary malignancies (n=3). Finally, 129 patients (73 men, 56 women) were included in the study.

Clinical characteristics

The clinical characteristics were retrieved from the electronic medical records, including age, sex, body mass index (BMI), total bilirubin (Tbil), albumin, CA19-9, carbohydrate antigen 125 (CA125) and carcinoembryonic antigen. Peripheral inflammatory markers were calculated as follows: neutrophil-to-lymphocyte ratio (NLR) = neutrophil count/lymphocyte count, platelet-to-white blood cell ratio (PWR) = platelet count/white blood cell count, platelet-to-lymphocyte ratio (PLR) = platelet count/lymphocyte count, lymphocyte-to-monocyte ratio (LMR) = lymphocyte count/monocyte count.

Histopathological analysis

Tumor grading was assessed on resected surgical specimens according to the 2019 World Health Organization classification system, which categorizes PDAC as well-differentiated, moderately differentiated, or poorly differentiated (23). In line with established methodologies (24,25), well- and moderately differentiated tumors were classified as low-grade (LG), whereas poorly-differentiated tumors were defined as high-grade (HG).

Spectral CT image acquisition

Imaging acquisition was conducted on a spectral CT scanner (IQon spectral CT, Philips Healthcare). The following scanning parameters were employed: tube voltage, 120 kVp; smart mAs; turn-around time, 0.5 s; detector layers, 64×0.625 mm; matrix, 512×512. Non-ionic contrast media (Ioversol, 350 mgI/mL) was injected into the median elbow vein at a rate of 3–4 mL/s using an automatic high–pressure injector, following the standard dose of 1.5 mL/kg body weight. Enhanced images of the arterial phase and portal vein phase (PVP) were acquired 12 s and 35 s after reaching a threshold level of 150 Hounsfield units (HUs) in the abdominal aorta (activated bolus tracking), respectively.

Spectral CT quantitative parameters analysis

The PVP spectral-based imaging data were transferred to the post-processing software [IntelliSpace Medicina Scientia (ISMS) Version 3.1.0, Philips Healthcare, China]. The software automatically generated IC image, monoenergetic images [40 kiloelectron volt (keV), 100 keV], and Zeff image. A 3D volume of interest (VOI) of the tumor was manually delineated layer by layer on the axial slices by two abdominal radiologists independently. Inter-observer and intra-observer reproducibility analyses were performed on 30 randomly selected patients. The VOI of 30 patients were re-delineated by the same radiologist after a month interval and by another radiologist. The software automatically generated the quantitative parameters, including 3D-40 keV, 3D-100 keV, 3D-Zeff, and 3D-IC. The slope of spectral HU curve (λHU) was calculated as: 3D-λHU = (3D-40 keV−3D-100 keV)/(100−40). The quantitative parameters extracted from these patients were subsequently assessed by the interclass correlation coefficient (ICC).

Variable selection, building and assessment of models

All 129 patients were randomly assigned to either a training cohort (n=90) or a testing cohort (n=39) at a ratio of 7:3. Univariate analysis was conducted to evaluate the variations in the clinical characteristics and 3D quantitative parameters between PDAC patients at LG and HG within the training cohort. To assess the robustness of the feature selection process, we additionally performed the least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation in the training cohort as a sensitivity analysis. The final identified variables were used to develop predictive models using three ML algorithms: LightGBM, MLP, and LR. A five-fold cross-validation method was performed on the training cohort. The testing cohort was used for independent validation. Model performance was evaluated by receiver operating characteristic (ROC) curves, area under the ROC curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, recall, and F1 score. Decision curve analysis (DCA) was employed to evaluate the clinical net benefit of the three models, and the corresponding confusion matrices for each classifier were presented for the testing cohort.

Statistical analysis

All statistical analyses were performed using Python (version 3.7.6) and SPSS software (version 25.0, IBM). The chi‑square test was used for categorical variables. Continuous variables were compared using the Mann-Whitney U test or the two-sample t-test, as appropriate. A two-tailed P<0.05 was considered statistically significant.

Results

Clinical characteristics and 3D quantitative parameters

Table 1 presents clinical characteristics and 3D quantitative parameters between PDAC patients in the training and testing cohorts. Table 2 presents clinical characteristics and 3D quantitative parameters for LG and HG PDAC patients in the training cohort. BMI, CA125, NLR, and all 3D quantitative parameters (3D‑conventional, 3D‑40 keV, 3D‑100 keV, 3D‑Zeff, 3D‑IC, and 3D‑λHU) differed significantly between LG and HG PDAC groups (P<0.05).

Variable selection, construction and validation of the ML models

Based on the results of univariate analysis, we performed LASSO regression with 10‑fold cross‑validation in the training cohort. At the optimal penalty parameter λ (λ =0.0152), which minimized the cross‑validated binomial deviance, LASSO identified nine variables with nonzero coefficients: BMI, CA125, NLR, 3D-conventional, 3D-40 keV, 3D-100 keV, 3D-Zeff, 3D-IC, and 3D-λHU (Table S1). These nine variables were used as inputs to develop three ML models, which were then independently validated in the testing cohort.

Evaluation of the three models

Table 3 summarizes the performance of the three models. Figure 1 shows the ROC curves for differentiating histopathological grades in the training and testing cohorts. The AUC of the LightGBM model in the training cohort was 0.872 [95% confidence interval (CI): 0.798–0.946], and in the testing cohort was 0.856 (95% CI: 0.670–1.000). The LR model achieved AUCs of 0.866 (95% CI: 0.790–0.943) in the training cohort and 0.859 (95% CI: 0.664–1.000) in the testing cohort, respectively. The MLP model achieved AUCs of 0.870 (95% CI: 0.788–0.953) in the training cohort and 0.796 (95% CI: 0.626–0.966) in the testing cohort, respectively. The accuracy of the three models in both cohorts is shown in Figure 2. The DCA demonstrated that all three models provided a positive net benefit in the testing cohort (Figure 3). Confusion matrices and calibration curves for the testing cohort are shown in Figures 4,5, respectively. Representative images are presented in Figure 6.

Figure 1 Performance of three ML models in the training and testing cohorts. AUC, area under the receiver operating characteristic curve; CI, confidence interval; LightGBM, light gradient boosting machine; LR, logistic regression; ML, machine learning; MLP, multilayer perceptron.

Figure 2 The accuracy of the three ML models in the training and testing cohorts. LightGBM, light gradient boosting machine; LR, logistic regression; ML, machine learning; MLP, multilayer perceptron classifier.

Figure 3 The DCA of three ML models in testing cohort. DCA, decision curve analysis; LightGBM, light gradient boosting machine; LR, logistic regression; ML, machine learning; MLP, multilayer perceptron classifier.

Figure 4 The confusion matrices of the three ML models in the testing cohort. LightGBM, light gradient boosting machine; LR, logistic regression; ML, machine learning; MLP, multilayer perceptron classifier.

Figure 5 The Calibration curves of three ML models in testing cohort. LightGBM, light gradient boosting machine; LR, logistic regression; ML, machine learning; MLP, multilayer perceptron classifier.

Figure 6 Representative multimodal imaging and pathological features of HG and LG PDAC. (A-H) Images of a 68-year-old patient (man) with a pathologically confirmed HG-PDAC. (A-C) The 40 keV imaging reveals a 3D-VOI (the light green area) from axial, coronal, and sagittal planes, and the CT value of 173.10 HU. (D) Pathologically confirmed as poorly differentiated PDAC (hematoxylin-eosin stain; original magnification, ×200). (E) The conventional image shows a 3D CT value of 80.93 HU. (F) The 100 keV image shows a 3D CT value of 56.72 HU. (G) The PVP iodine map reveals a 3D IC value of 1.56 mg/mL. (H) The Zeff map shows 3D Zeff value of 8.17. (I-P) Images of a 55-year-old patient (man) with a pathologically confirmed LG-PDAC. (I-K) The 40 keV imaging reveals a 3D-VOI (the light green area) from axial, coronal, and sagittal planes, and CT value of 97.37 HU. (L) Pathologically confirmed as moderately differentiated PDAC (hematoxylin-eosin stain; original magnification, ×100). (M) The conventional image shows a 3D VOI CT value of 47.87 HU. (N) The 100 keV image shows a 3D CT value of 42.51 HU. (O) The iodine map reveals a 3D IC value of 0.74 mg/mL. (P) The Zeff map shows 3D Zeff value of 7.72. 3D, three-dimensional; CT, computed tomography; HG, high-grade; HU, Hounsfield unit; IC, iodine concentration; keV, kiloelectron volt; LG, low-grade; PDAC, pancreatic ductal adenocarcinoma; PVP, portal venous phase; VOI, volume of interest; Zeff, effective atomic number.

Discussion

In this study, we developed three ML models incorporating clinical characteristics and 3D quantitative parameters to predict histopathological differentiation in PDAC. All models demonstrated good discriminative performance, and DCA indicated satisfactory net benefit. These results indicate that ML models can serve as a robust, noninvasive tool for identifying PDAC patients with HG pathology before surgery.

Studies indicate that contrast agents may not rapidly and completely fill the microvasculature with sufficient volume during the arterial phase (26). In contrast, the venous phase provides a longer delay after contrast injection, allowing adequate time for the agent to fill the microvasculature and leakage into the cellular interstitium (27). Hence, venousphase quantitative data more accurately reflect tumor perfusion. Several studies have shown that portal venous phase imaging holds significant potential for revealing the biological characteristics of PDAC (28,29). Because individual differences can substantially affect quantitative parameters (30), we adopted 3D quantitative analysis to more accurately evaluate the relationship between spectral CT parameters and histopathological differentiation. Unlike 2D analysis, which captures only incomplete tumor regions, 3D analysis encompasses the entire tumor volume. This comprehensive approach provides a more reliable evaluation of tumor characteristics and minimizes observer-dependent variability (31). Therefore, we focused exclusively on the portal venous phase to derive all 3D quantitative parameters and explored the diagnostic value of single‑phase quantitative parameters in a targeted manner.

In our study, 3D quantitative parameters—including 3D‑conventional, 3D‑40 keV, 3D‑100 keV, 3D‑Zeff, 3D‑IC, and 3DλHU—were substantially lower in the HG PDAC group than in the LG group. These parameters are simple to obtain and clinically practical. The internal structure of PDAC is highly complex, with perfusion characteristics influenced by fibrosis content, cellular density, residual normal pancreatic cells, and blood supply. In LG PDAC, microvascular density within the pancreatic acini is relatively higher, leading to a richer blood supply (32). Additionally, the stromal compartment surrounding HG PDAC cells is denser and characterized by increased collagen deposition (33). Collectively, these factors may explain the slightly higher quantitative parameters in LG PDAC compared with HG PDAC, consistent with previous findings (10).

Our study also found that BMI, CA125, and NLR, were identified as clinical characteristics for differentiating between HG and LG PDAC. The association between reduced BMI and HG PDAC is rooted in a bidirectional interplay. Tumor-induced hypercatabolic states drive systemic metabolic consumption, leading to cachexia and weight loss. Obesity-associated metabolic dysregulation—characterized by chronic inflammation and insulin resistance—can promote malignant progression and more aggressive phenotypes (34,35). Elevated CA125 levels are a recognized biomarker associated with high tumor burden, aggressive biology, and metastatic potential in PDAC, and thus correlates with poorer prognosis (36,37). Therefore, the HG PDAC is associated with CA125 levels. The pro-inflammatory microenvironment promotes carcinogenesis in pancreatic tissue, and malignant cells reciprocally amplify inflammatory responses through cytokine hypersecretion and immune cell recruitment, establishing a self-perpetuating cycle (38,39). Several studies have shown that quantitative and functional alterations in neutrophils and lymphocytes—particularly as reflected by an elevated NLR—are mechanistically linked to tumor initiation, progression, and metastatic (40,41).

ML, a subset of artificial intelligence, offers excellent data processing capabilities and can capture complex correlations among variables. Recently, ML applications in PDAC have primarily focused on differential diagnoses, assessment of lymph node metastasis, and prediction of survival and recurrence risk (42-44). All ML models developed by Xu et al. (45) demonstrated strong performance in predicting histological grading in PDAC. Consistent with previous studies, our research validates the potential of ML for the accurate, non-invasive assessment of histopathological grade. In the preoperative setting for PDAC, where treatment decisions hinge on tumor grade, an AUC below 0.80 is typically considered of limited clinical utility (46). Our models achieved testing AUCs of 0.856 (LightGBM), 0.859 (LR), and 0.796 (MLP)—with the first two exceeding the 0.80 threshold—and training AUCs ranging from 0.856 to 0.872. By integrating spectral CT‑derived 3D quantitative parameters with three ML algorithms, this study presents a novel, computationally efficient tool for noninvasive PDAC grading, with all models exhibiting satisfactory discrimination. Although calibration was suboptimal (Brier scores: 0.227–0.309), the DCA confirmed that all models offered positive net benefit in the testing cohort, highlighting their clinical applicability. However, the LightGBM model’s low sensitivity (0.444) in the testing cohort—missing more than half of HG PDAC patients—limits its clinical utility as a preoperative screening tool. In contrast, the LR model, with a markedly higher sensitivity (0.769), appears more clinically relevant for identifying candidates with HG disease. As this is a preliminary exploratory analysis, the optimal model choice should be guided by the specific clinical objective.

Limitations

There are certain limitations in this study. First, its single‑center design and relatively small sample size—particularly the small number of HG PDACs—increase the risk of overfitting despite internal validation, while the consistency between univariate and LASSO feature selection supports the stability of the predictor set, rendering our findings exploratory. Additionally, the retrospective design may introduce inherent bias, and validation on a single‑center dataset restricts generalizability. Therefore, large‑scale, multicenter prospective studies are needed before clinical implementation.

Conclusions

We developed three ML models integrating 3D quantitative parameters and clinical features for preoperative prediction of PDAC tumor grade. All models exhibited robust accuracy, validating 3D quantitative parameters from spectral CT as reliable biomarkers for assessing histopathological grade and highlighting their clinical utility.

Acknowledgments

The authors thank all volunteers who participated in the study and the staff of the Department of Radiology, Chongqing General Hospital, China, for their selfless and valuable assistance.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2025-1-1023/rc

Data Sharing Statement: Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2025-1-1023/dss

Peer Review File: Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2025-1-1023/prf

Funding: This study was supported by the Medical Research Youth Program of the Chongqing National Health Commission and Chongqing Science and Technology Bureau, China (No. 2024QNXM058), the Key Special Program of Technological Innovation and Application Development in Chongqing, China (No. CSTB2023TIAD-KPX0059-2), and the General Program of Natural Science Foundation of Chongqing, China (No. CSTB2025NSCQ-GPX0311).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2025-1-1023/coif). X.Z. is employed by the company Philips Healthcare. The other 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. This study was approved by the Institutional Review Board of the Chongqing General Hospital (No. KY S2024-069-01), individual consent for this retrospective analysis was waived.

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. Siegel RL, Miller KD, Wagle NS, et al. Cancer statistics, 2023. CA Cancer J Clin 2023;73:17-48. [Crossref] [PubMed]
2. Xia C, Dong X, Li H, et al. Cancer statistics in China and United States, 2022: profiles, trends, and determinants. Chin Med J (Engl) 2022;135:584-90. [Crossref] [PubMed]
3. Brown ZJ, Cloyd JM. Surgery for pancreatic cancer: recent progress and future directions. Hepatobiliary Surg Nutr 2021;10:376-8. [Crossref] [PubMed]
4. Zhang XP, Xu S, Gao YX, et al. Early and late recurrence patterns of pancreatic ductal adenocarcinoma after pancreaticoduodenectomy: a multicenter study. Int J Surg 2023;109:785-93. [Crossref] [PubMed]
5. McKinney M, Griffin MO, Tolat PP. Multimodality Imaging for the Staging of Pancreatic Cancer. Surg Oncol Clin N Am 2021;30:621-37. [Crossref] [PubMed]
6. Bhatti I, Ojo D, Dennison AR, et al. Percutaneous Pancreatic Biopsies-Still an Effective Method for Histologic Confirmation of Malignancy. Surg Laparosc Endosc Percutan Tech 2016;26:334-7. [Crossref] [PubMed]
7. Gerlinger M, Rowan AJ, Horswell S, et al. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. N Engl J Med 2012;366:883-92. [Crossref] [PubMed]
8. Poruk KE, Gay DZ, Brown K, et al. The clinical utility of CA 19-9 in pancreatic adenocarcinoma: diagnostic and prognostic updates. Curr Mol Med 2013;13:340-51. [Crossref] [PubMed]
9. Wang SH, Sun YF, Liu Y, et al. CT contrast enhancement correlates with pathological grade and microvessel density of pancreatic cancer tissues. Int J Clin Exp Pathol 2015;8:5443-9.
10. Liu W, Xie T, Chen L, et al. Dual-layer spectral detector CT: A noninvasive preoperative tool for predicting histopathological differentiation in pancreatic ductal adenocarcinoma. Eur J Radiol 2024;173:111327. [Crossref] [PubMed]
11. McCollough CH, Leng S, Yu L, et al. Dual- and Multi-Energy CT: Principles, Technical Approaches, and Clinical Applications. Radiology 2015;276:637-53. [Crossref] [PubMed]
12. Wu L, Cen C, Yue X, et al. A clinical-radiomics nomogram based on dual-layer spectral detector CT to predict cancer stage in pancreatic ductal adenocarcinoma. Cancer Imaging 2024;24:55. [Crossref] [PubMed]
13. Doi S, Yanagawa M, Matsui T, et al. Usefulness of Three-Dimensional Iodine Mapping Quantified by Dual-Energy CT for Differentiating Thymic Epithelial Tumors. J Clin Med 2023;12:5610. [Crossref] [PubMed]
14. Gierada DS, Rydzak CE, Zei M, et al. Improved Interobserver Agreement on Lung-RADS Classification of Solid Nodules Using Semiautomated CT Volumetry. Radiology 2020;297:675-84. [Crossref] [PubMed]
15. Zeng D, Zhang J, Song Z, et al. Development and validation of a model based on preoperative dual-layer detector spectral computed tomography 3D VOI-based quantitative parameters to predict high Ki-67 proliferation index in pancreatic ductal adenocarcinoma. Insights Imaging 2024;15:291. [Crossref] [PubMed]
16. Goecks J, Jalili V, Heiser LM, et al. How Machine Learning Will Transform Biomedicine. Cell 2020;181:92-101. [Crossref] [PubMed]
17. Liu L, Si M, Ma H, et al. A hierarchical opportunistic screening model for osteoporosis using machine learning applied to clinical data and CT images. BMC Bioinformatics 2022;23:63. [Crossref] [PubMed]
18. Li Y, Wu Y, Gao Y, et al. Machine-learning based prediction of prognostic risk factors in patients with invasive candidiasis infection and bacterial bloodstream infection: a singled centered retrospective study. BMC Infect Dis 2022;22:150. [Crossref] [PubMed]
19. Nakanishi R, Slomka PJ, Rios R, et al. Machine Learning Adds to Clinical and CAC Assessments in Predicting 10-Year CHD and CVD Deaths. JACC Cardiovasc Imaging 2021;14:615-25. [Crossref] [PubMed]
20. Ben Hmido S, Abder Rahim H, Ingwersen EW, et al. Transparency in Externally Validated Models: A systematic review of machine learning vs. logistic regression for predicting colorectal anastomotic leakage. Eur J Surg Oncol 2025;51:110367.
21. Wang S, Tang Y, Zou P, et al. Quantitative LIBS Analysis of Microalloyed Steel Using LASSO-LightGBM Modeling with SHAP-Based Interpretability. J Anal At Spectrom 2025;40:3563-8.
22. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436-44. [Crossref] [PubMed]
23. Nagtegaal ID, Odze RD, Klimstra D, et al. The 2019 WHO classification of tumours of the digestive system. Histopathology 2020;76:182-8. [Crossref] [PubMed]
24. Qiu W, Duan N, Chen X, et al. Pancreatic Ductal Adenocarcinoma: Machine Learning-Based Quantitative Computed Tomography Texture Analysis For Prediction Of Histopathological Grade. Cancer Manag Res 2019;11:9253-64. [Crossref] [PubMed]
25. Tikhonova VS, Karmazanovsky GG, Kondratyev EV, et al. Radiomics model-based algorithm for preoperative prediction of pancreatic ductal adenocarcinoma grade. Eur Radiol 2023;33:1152-61. [Crossref] [PubMed]
26. Zhang Z, Zou H, Yuan A, et al. A Single Enhanced Dual-Energy CT Scan May Distinguish Lung Squamous Cell Carcinoma From Adenocarcinoma During the Venous phase. Acad Radiol 2020;27:624-9. [Crossref] [PubMed]
27. Tian S, Jianguo X, Tian W, et al. Application of dual-energy computed tomography in preoperative evaluation of Ki-67 expression levels in solid non-small cell lung cancer. Medicine (Baltimore) 2022;101:e29444. [Crossref] [PubMed]
28. Kim BR, Kim JH, Ahn SJ, et al. CT prediction of resectability and prognosis in patients with pancreatic ductal adenocarcinoma after neoadjuvant treatment using image findings and texture analysis. Eur Radiol 2019;29:362-72. [Crossref] [PubMed]
29. Attiyeh MA, Chakraborty J, McIntyre CA, et al. CT radiomics associations with genotype and stromal content in pancreatic ductal adenocarcinoma. Abdom Radiol (NY) 2019;44:3148-57. [Crossref] [PubMed]
30. Zopfs D, Reimer RP, Sonnabend K, et al. Intraindividual Consistency of Iodine Concentration in Dual-Energy Computed Tomography of the Chest and Abdomen. Invest Radiol 2021;56:181-7. [Crossref] [PubMed]
31. Zhong LJ, Yu N, Zhou XJ, et al. Differentiating between pulmonary adenocarcinoma and squamous cell carcinoma by spectral CT volumetric quantitative analysis: a comparative study with conventional spectral analysis. J Thorac Dis 2023;15:679-89. [Crossref] [PubMed]
32. Yoon SH, Lee JM, Cho JY, et al. Small (≤ 20 mm) pancreatic adenocarcinomas: analysis of enhancement patterns and secondary signs with multiphasic multidetector CT. Radiology 2011;259:442-52. [Crossref] [PubMed]
33. Laklai H, Miroshnikova YA, Pickup MW, et al. Genotype tunes pancreatic ductal adenocarcinoma tissue tension to induce matricellular fibrosis and tumor progression. Nat Med 2016;22:497-505. [Crossref] [PubMed]
34. Tong J, Jiang W, Mao S, et al. Development and validation of a nomogram to predict liver metastasis for pancreatic ductal adenocarcinoma after radical resection. Front Oncol 2022;12:1040411. [Crossref] [PubMed]
35. Liu L, Xiang J, Chen R, et al. The clinical utility of CA125/MUC16 in pancreatic cancer: A consensus of diagnostic, prognostic and predictive updates by the Chinese Study Group for Pancreatic Cancer (CSPAC). Int J Oncol 2016;48:900-7. [Crossref] [PubMed]
36. Carreras-Torres R, Johansson M, Gaborieau V, et al. The Role of Obesity, Type 2 Diabetes, and Metabolic Factors in Pancreatic Cancer: A Mendelian Randomization Study. J Natl Cancer Inst 2017;109:djx012. [Crossref] [PubMed]
37. Michaud DS, Giovannucci E, Willett WC, et al. Physical activity, obesity, height, and the risk of pancreatic cancer. JAMA 2001;286:921-9. [Crossref] [PubMed]
38. Elshaer M, Gravante G, Kosmin M, et al. A systematic review of the prognostic value of lymph node ratio, number of positive nodes and total nodes examined in pancreatic ductal adenocarcinoma. Ann R Coll Surg Engl 2017;99:101-6. [Crossref] [PubMed]
39. Singh N, Baby D, Rajguru JP, et al. Inflammation and cancer. Ann Afr Med 2019;18:121-6. [Crossref] [PubMed]
40. Xiang ZJ, Hu T, Wang Y, et al. Neutrophil-lymphocyte ratio (NLR) was associated with prognosis and immunomodulatory in patients with pancreatic ductal adenocarcinoma (PDAC). Biosci Rep 2020;40:BSR20201190. [Crossref] [PubMed]
41. Iwai N, Okuda T, Sakagami J, et al. Neutrophil to lymphocyte ratio predicts prognosis in unresectable pancreatic cancer. Sci Rep 2020;10:18758. [Crossref] [PubMed]
42. Zhang L, Chen X, Chen Z, et al. Differentiating pancreatic ductal adenocarcinoma and autoimmune pancreatitis using a machine learning model based on ultrasound clinical features. Front Oncol 2025;15:1505376. [Crossref] [PubMed]
43. Qu C, Zeng P, Li C, et al. A machine learning model based on preoperative multiparametric quantitative DWI can effectively predict the survival and recurrence risk of pancreatic ductal adenocarcinoma. Insights Imaging 2025;16:38. [Crossref] [PubMed]
44. Daibo S, Homma Y, Ohya H, et al. Novel machine-learning model for predicting lymph node metastasis in resectable pancreatic ductal adenocarcinoma. Ann Gastroenterol Surg 2025;9:161-8. [Crossref] [PubMed]
45. Xu Y, Shi Y, Jiang T, et al. Radiomics-based histological grading of pancreatic ductal adenocarcinoma using (18)F-FDG PET/CT: A two-center study. Eur J Radiol 2025;187:112070. [Crossref] [PubMed]
46. Lekkas G, Vrochidou E, Papakostas GA. Advancements in Radiomics-Based AI for Pancreatic Ductal Adenocarcinoma. Bioengineering (Basel) 2025;12:849. [Crossref] [PubMed]