All Times ET
Keywords: CT scan, dose reduction, image prediction, hepatic artery, portal vein
Multiphase CT recommended by American Association for the Study of Liver Diseases (AASLD) in liver disease diagnoses helps early detection of hepatocellular carcinoma (HCC), the primary liver cancer, and directly benefits patient care and survival rate. A normal multiphase CT usually consists of four CT scans, one scan without contrast agent as the baseline and three scans at different times with intravenous-injected contrast agents. The four scans help the differentiation of normal and tumor tissues with enhanced contrast difference in CT images. While a variety of studies have been conducted to reduce the dose of CT scans and the risks of contrast agents, a patient with a multiphase CT scan still receives a high radiation dose. The objective of this study is to derive the delayed phase scan in the multiphase CT protocol from the other three scans using machine learning, which will reduce 25% of the radiation dose. A revised U-Net architecture was trained by manually segmented and intra-subject registered liver images with preprocessed hepatic artery phase and portal vein phase images after removing the baseline as inputs and delayed phase images after subtracting the baseline as targets. Data argumentation methods were implemented to increase the training datasets. The trained deep learning network predicts the difference between the output volume and the baseline volume. The final delayed phase images then were calculated by combining the predicted offset and the baseline images. Results show that the deep neural network can predict the delayed phase scan images in high accuracy with a mean squared error (MSE) of 0.0004 for images of 512 by 512 pixels. The ultimate goal of this research is to reduce two phase volumes for 50% radiation dose reduction. The portal vein scan and the delayed phase scan will be derived from the baseline and hepatic artery phase scans.