Unsupervised deep learning based image outpainting for dual-source, dual-energy computed tomography

Abstract

Due to a physical constraint, dual-energy computed tomography (DECT) performed on a dual-source CT scanner has a restricted field of view (FOV) in one imaging chain (i.e. tube B, 140 kV with tin filter). This indicates that dual-energy analysis cannot be performed outside the limited FOV. As a result, the dual-source DECT scanner may not be beneficial for larger patients. To address this issue, we study the feasibility of using an unsupervised deep learning (DL) based method to outpaint the missing image outside the limited FOV. Brain DECT images acquired on the dual-source DECT scanner were used to simulate a restricted FOV scan. First, the whole brain of DECT images was shifted to the corner of the image. Then, the restricted 140-kV CT image was generated by multiplying the shifted 140-kV CT image with a small circular mask. As a result, we can evaluate the proposed DL-based method without scanning the patient twice. Moreover, the non-truncated 140-kV CT images can be considered as ground truth for comparison. Our results show that the proposed DL-based method can reconstruct missing data and produce 140-kV CT images that appear similar to the true 140-kV CT images. Moreover, the mean CT number differences between the true and DL-based 140-kV CT images for brain, muscle, fat and bone were less than 3 HU. We also observed that virtual monoenergetic images obtained from true and DL-based DECT images were visually similar. Our preliminary study shows the feasibility of using an unsupervised DL-based method to yield the out-of-field imaging data in dual-source DECT.