Deformable Lung CT Registration by Decomposing Large Deformation

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.