Deformable Lung CT Registration by Decomposing the Large Deformation

Jing Zou, Lihao Liu, Youyi Song, Kup-Sze Choi, Jing Qin

Keywords: Image registration · Lung CT · Organ movement · Deformation field decomposition · Attention layer.

TL;DR: A patient-specific lung CT registration method by decomposing the large deformation field into small fields and refining them with an attention layer.

Abstract: Deformable lung CT registration plays an important role in image-guided navigation systems, especially in the situation with organ motion. Recent progress has been made in image registration by utilizing neural networks for end-to-end inference of a deformation field. However, there are still difficulties to learn the irregular and large deformation caused by organ motion. In this paper, we propose a patient-specific lung CT image registration method. We first decompose the large deformation between the source image and the target image into several continuous intermediate fields. Then we compose these fields to form a spatio-temporal motion field and refine it through an attention layer by aggregating information along motion trajectories. The proposed method can utilize the temporal information in a respiratory circle and can generate intermediate images which are helpful in image-guided systems for tumor tracking. Extensive experiments were performed on a public dataset, showing the validity of the proposed methods.