Can Machine Learning Predict Metastatic Sites in Pancreatic Ductal Adenocarcinoma? A Radiomic Analysis

Area of Science:

• Oncology
• Radiology
• Medical Imaging

Background:

• Pancreatic ductal adenocarcinoma (PDAC) has a high potential for metastasis, significantly impacting patient prognosis.
• Metastatic sites in PDAC, such as the liver or lungs, are associated with distinct clinical outcomes.
• Radiomics offers quantitative analysis of medical images for predictive modeling in cancer.

Purpose of the Study:

• To assess the feasibility of using radiomic models to predict the metastatic patterns of PDAC.
• To differentiate between hepatic and pulmonary metastases from primary pancreatic tumors using radiomic features.

Main Methods:

• Retrospective analysis of CT scans from 115 PDAC patients with liver or lung metastases.
• Extraction of radiomic features from primary tumors using PyRadiomics software.
• Feature selection via LASSO regularization, with class imbalance addressed by SMOTE and class weighting.

Main Results:

• A multivariate logistic regression model achieved an AUC-ROC of 0.831.
• The model demonstrated a sensitivity of 0.762 and specificity of 0.787 for distinguishing metastatic patterns.
• A high negative predictive value (0.810) for lung metastases was observed, with 'LargeDependenceEmphasis' showing a trend as a discriminative feature.

Conclusions:

• Radiomic analysis of primary pancreatic tumors shows promise in predicting hepatic versus pulmonary metastatic patterns.
• The model's high negative predictive value for lung metastases may aid clinical decision-making.
• External validation in larger cohorts is necessary for clinical implementation.