Pancreatic Cancer, Radiomics and Artificial Intelligence: A Review

Abstract

Computed tomography (CT) scans stratified patients with pancreatic ductal adenocarcinoma (PDA) into categories based on whether the tumor is expected to be resectable, borderline resectable, initially unresectable, or metastatic. When reporting these exams, radiologists use structured templates to ensure that the generated information is complete, although the difficulty in identifying initial microscopic infiltrations of adjacent structures and small metastases is well recognized. Radiomics is seen as a potentially useful tool for determining tumor aggressiveness and building predictive clinical models. If extracted radiomic signatures are validated as prognostic and predictive biomarkers, they could be used aiding in decision-making to facilitate personalized patient management with ACDP. Models with convolutional neural networks provide estimations associated with a biological aggressiveness profile by combining clinical, semantic, and radiomic features. Despite encouraging results, the main limitations for clinical use of quantitative imaging are due to the instability of the measurements and the diversity of obtained images (different equipment and protocols), both making difficult to generalize the obtained results. The availability of large multicenter repositories with standardized and annotated images, and associated data (clinical, molecular, genetic), together with radiomics and artificial intelligence tools, will allow to predict the behavior of these tumors at the diagnosis. Its validation in totally independent cohorts and causal inference models is needed.