Quantitative CT-based radiomics of pancreatic ductal adenocarcinoma: a valuable tool for radiological staging?

Purpose: To investigate the value of quantitative radiomics to predict local resectability and presence of metastatic disease of pancreatic ductal adeno-carcinoma (PDAC) on routine contrast-enhanced abdominal CT.Material and methods: 89 patients diagnosed with stage I-IV PDAC in the head were included (m:f=47:42; range 45-84 yrs). First-order intensity, texture and shape features were calculated for manually 3D-segmented tumors. Segmenta-tion of local vessels was not performed. Student’s t test or the Mann-Whitney test were used to compare both locally advanced (n=36, 40%) and/or metastatic tumors (n=18, 20%) to resected tumors (n=42, 47%). A feature was considered significant with p<0.05. Using these features, two support vector machines (SVM) were trained: one to predict local resectability and one to predict the presence of distant metastases.Results: Following an inter-observer analysis in eleven cases, 84/105 features (ICC >0.75) were selected for further analysis. Univariate analysis resulted in 24 features associated with locally advanced disease, and 10 features associated with metastatic disease. These features were used to predict local resectability and distant metastases using a SVM. For local resectability, this resulted in a sensitivity and specificity of 75% and 67%, with an area under curve (AUC) of 0.71, comparable to experts sensitivity and specificity of 60-96% and 22-76%. The prediction of metastatic disease resulted in a sensitivity and specificity of 40% and 60%, with an AUC of 0.51.Conclusion: Quantitative CECT-based radiomics, based on tumor features only, is comparable to expert performance in predicting local resectability. Prediction of distant metastases is more challenging.