Predicting Synchronous Liver Metastasis in Pancreatic Cancer Using CT Radiomics and Clinical Features: A Machine Learning Approach.
TL;DR
The integrated model enables accurate preoperative prediction of synchronous LM in PC and is designed to assist clinical decision-making in the common dilemma of indeterminate subcentimeter liver lesions on CT: a low-risk score supports surveillance and spares unnecessary biopsy, whereas a high-risk score prompts confirmatory biopsy or a shift toward neoadjuvant systemic therapy.
Abstract
Rationale
AND
Objectives
To address the challenge of preoperative prediction of synchronous liver metastasis (LM) in pancreatic cancer (PC), we developed and validated machine learning models integrating clinical and computed tomography (CT) radiomics features, and compared the performance and interpretability of linear (linear discriminant analysis [LDA]) versus nonlinear (multilayer perceptron [MLP]) architectures.
Materials And Methods
This retrospective study enrolled 340 patients with pancreatic ductal adenocarcinoma (190 with LM, 150 without). Five radiomics features were selected using the minimum redundancy maximum relevance and least absolute shrinkage and selection operator methods. Based on independent clinical predictors (age, carcinoembryonic antigen [CEA], and clinical N stage) and the selected radiomics features, 26 machine learning models were constructed. The optimal models (LDA and MLP) were evaluated in an independent validation cohort (n = 102) using receiver operating characteristic curves, calibration, decision curve analysis, and SHAP (SHapley Additive exPlanations) analysis.
Results
In the validation cohort, LDA achieved an area under the curve (AUC) of 0.828 (95% confidence interval [CI]: 0.735-0.897), and MLP achieved 0.822 (95% CI: 0.732-0.895). Both models demonstrated good calibration, with Hosmer-Lemeshow test P values of 0.551 (Brier score = 0.168) for LDA and 0.682 (Brier score = 0.172) for MLP, alongside high net clinical benefit. SHAP analysis revealed that clinical N stage dominated the LDA model, whereas CEA and radiomics features played more prominent roles in the MLP model. Friedman's H-statistic identified five significant feature interactions. Subgroup analysis showed that MLP maintained stable performance (AUC 0.66-0.77), whereas LDA performed poorly in the lymph-node-negative (AUC = 0.50) and pancreatic head tumor (AUC = 0.62) subgroups. The clinical nomogram achieved an AUC of 0.788 in the validation cohort.
Conclusion
The integrated model enables accurate preoperative prediction of synchronous LM in PC. Rather than replacing pathological diagnosis, this tool is designed to assist clinical decision-making in the common dilemma of indeterminate subcentimeter liver lesions on CT: a low-risk score supports surveillance and spares unnecessary biopsy, whereas a high-risk score prompts confirmatory biopsy or a shift toward neoadjuvant systemic therapy. Both architectures exhibit robust discrimination and clinical utility; however, the nonlinear MLP model shows superior stability across clinical subgroups, offering a promising tool for individualized risk assessment and treatment planning.