Self-supervised deformable registration of digital tomosynthesis and 3D CT images for surgical navigation

To enhance the precision and efficiency of surgical procedures, there is an increasing demand for integrating medical imaging systems into the surgical environment. For example, a mobile C-Arm image has been widely applied to intraoperative imaging system. However, the projective nature of C-Arm image limits its effectiveness for accurate surgical guidance. Recently, a mobile digital tomosynthesis (DTS) system has emerged as an attractive alternative, providing quasi-3D images with a radiation dose comparable to that of conventional 2D radiography and ease of use in the intraoperative scene. To enhance the utility of mobile DTS, it is required to register DTS with preoperative CT. However, the previously proposed registration methods have shown insufficient results due to the limited depth resolution and the streaking artifact inherent to DTS. Therefore, we propose a novel self-supervised learning-based 3D deformable registration between CT and mobile DTS by introducing a dataset generation technique to train the deep learning network with an additional constraint. The proposed registration algorithm is composed of a sequential process involving an affine registration based on iterative optimization and a deformable registration based on self-supervised learning. To evaluate the registration accuracy, we compared the distances between matching points. We observed that the 3D distance decreased from 14.07±1.79 voxels to 3.73±0.53 voxels after affine registration and to 2.55±0.50 voxels after the self-supervised deformable registration.