Unsupervised/Supervised Hybrid Deep Learning Framework for Low Dose Phase Contrast Imaging

Abstract

X-ray phase contrast computed tomography (PCCT) has better imaging quality than conventional attenuation X-ray CT and has demonstrated promising application prospects in medical diagnosis. However, reducing the radiation dose during PCCT imaging still remains a major challenge. Recently, deep learning (DL) techniques have been applied to low dose CT and obtain significant progress. Most of them require massive paired images to train the network in a supervised manner, which may hamper their practical applications because the ground-truth images are hard to be obtained in most cases. To address this issue, we report a hybrid deep learning framework for low dose PCCT which capsules unsupervised and supervised learning manners. It combines the advantages of convolutional neural network (CNN) and total variation (TV) and is suitable for both unlabelled datasets and labelled datasets. This framework has been validated and demonstrated with experimental data. It will be helpful to push the practical application of low dose PCCT.