Automatic Liver Segmentation in Pre-TIPS Cirrhotic Patients: A Preliminary Step for Radiomics Studies

The aim of this study is to present a deep learning (DL) algorithm for accurate liver delineation in high-resolution computed tomography (CT) images of pre-transjugular intrahepatic portosystemic shunt (TIPS) cirrhotic patients. In this way, we aim to improve the methodology performed by medical physicians in radiomics studies where the use of operator-independent segmentation methods is mandatory to correctly identify the target and to obtain accurate predictive models. Two DL models were investigated: UNet, the most widely used DL network for biomedical image segmentation, and the innovative customized efficient neural network (C-ENet). 111 patients with liver contrast-enhanced CT examinations before TIPS procedure were considered. The performance of the two DL networks was evaluated in terms of the similarity of their segmentations to the gold standard. The results show that C-ENet can be used to obtain accurate (dice similarity coefficient = 87.70%) segmentation of the liver region outperforming UNet (dice similarity coefficient = 85.33%). In conclusion, we demonstrated that DL can be efficiently applied to rapidly segment cirrhotic liver images, without any radiologist supervision, to produce user-independent results useful for subsequent radiomics studies.