X-CTRSNet: 3D cervical vertebra CT reconstruction and segmentation directly from 2D X-ray images

Abstract

Orthogonal 2D cervical vertebra (C-vertebra) X-ray images have the advantages of high imaging efficiency, low radiation risk, easy operation and low cost for rapid primary clinical diagnoses. Especially in emergency departments, this technique is known to be significantly useful in triage for trauma patients. However, the technique can only provide overlapping anatomic information from limited projection views and is unable to visually exhibit full-view anatomy and precise stereo structures without further CT examination. To promote “once is enough” for visualizing 3D anatomy & structures and reducing repetitive radiation as much as possible, we proposed X-CTRSNet for 2D X-ray images. This is the first powerful work that simultaneously and accurately enables 3D C-vertebra CT reconstruction and segmentation directly from orthogonally anteroposterior- and lateral-view 2D X-ray images. X-CTRSNet combines the reciprocally coupled SpaDRNet for reconstruction & MulSISNet for segmentation, and a RSC Learning for tasks consistency. The experiment shows that X-CTRSNet successfully reconstructs and segments the 3D C-vertebra CT from the 2D X-ray images with a PSNR of 24.58 dB, an SSIM of 0.749, and an average Dice of 80.44%. All these findings reveal the great potential of X-CTRSNet in clinical imaging and diagnosis to facilitate emergency triage by enabling precise 3D reconstruction and segmentation on 2D X-ray images.