Low dose 4D-CT super-resolution reconstruction via inter-plane motion estimation based on optical flow

摘要

In order to obtain the dynamic intuitive image of the patient's internal organs movement and minimize the potential risks of X-ray radiation at the same time, low dose four-dimensional computed tomography (4D-CT) has attracted a considerable interest in the high precision radiation therapy. But some susceptible artifacts, including device-dependent, image reconstruction times and patient's respiratory pattern, usually cause an inter-plane thickness that is much greater than intra-plane voxel resolutions. In this study, to estimate the respiratory motion and enhance the inter-plane resolution of multi-plane computed tomography (CT) images, a joint optimization framework was proposed using the combined local and global (CLG) variational optical flow and improved non-local iterative back projection (NLIBP). Note that, the premise of this work is that the anatomical information missing in one particular phase can be recovered from other phases in CT images. First, CLG variational optical flow model was constructed to estimate the respiratory motion (i.e., the optical flow fields) between different phases at the corresponding voxel positions, and then was solved by the fast alternating direction method of multipliers (ADMM). Secondly, the improved NLIBP algorithm characterized by non-local mean filter and image fusion strategy was employed to reconstruct high resolution (HR) inter-plane images based on the calculated motion fields. Finally, we explored different hyperparameter settings to achieve a good trade-off between the super resolution (SR) reconstruction performance and computational efficiency, and indicated the success of CLG variational optical flow method for estimating the displacement field between images. Experimental results on public lung 4D-CT datasets demonstrated that this proposed method is able to more effectively enhance texture structures while preserving edges, and outperforms current state-of-the-art methods both quantitatively and qualitatively. (C) 2020 Elsevier Ltd. All rights reserved.