Radiomics in differential diagnosis of pancreatic tumors

Abstract

The aim of this study was to assess whether radiomics could predict histotype of pancreatic ductal adenocarcinomas (PDAC) and pancreatic neuroendocrine tumors (PNET). Contrast-enhanced CT scans of 193 patients were retrospectively reviewed, encompassing 97 PDACs and 96 PNETs. Additionally, anamnestic data and laboratory data were evaluated. A total of 107 features were extracted for both the arterial and venous phases. ROC curves were constructed for the parameters with the highest AUC, considering two groups: one including all lesions and the other including only lesions smaller than 5 cm. The following feature differences were found to be statistically significant (p < 0.05). Without considering lesion size: for the arterial phase, 16 first-order and 38 s-order features; for the venous phase, 10 first-order and 20 s-order features. When considering lesion size: for the arterial phase, 16 first-order and 52 s-order features; for the venous phase, 11 first-order and 36 s-order features. The radiomics features with the highest AUC values included ART_firstorder_RootMeanSquared (AUC = 0.896, p < 0.01) in the arterial phase and VEN_firstorder_Median (AUC = 0.737, p < 0.05) in the venous phase for all lesions, and ART_firstorder_RootMeanSquared (AUC = 0.859, p < 0.01) and VEN_firstorder_Median (AUC = 0.713, p < 0.05) for lesions smaller than 5 cm. Texture analysis of pancreatic pathology has shown good predictability in defining the PNET histotype. This analysis potentially offering a non-invasive, imaging-based method to accurately differentiate between pancreatic tumor types. Such advancements could lead to more precise and personalized treatment planning, ultimately optimizing the use of medical resources.