A U-Shaped Convolution Recurrent Network for Spatiotemporal Feature Fusion in 4D-CT Image Registration

In lung 4DCT registration, pairwise methods do not fully utilize the temporal component of serial images. Recently some studies have tried to address this issue. However, interpretability and high-dimensional feature modeling in a temporally explicit manner still need to be further explored. In this paper, a U-shaped network based on Conv-LSTM was proposed that defines image registration as a spatiotemporal sequence prediction problem. The backbone structure consists of stacked Conv-LSTM layers, and a time channel is introduced to fuse the spatio-temporal features of image sequences. We train model in one-shot style on DIRLAB dataset. The experimental results demonstrate that the proposed method achieves better registration performance in spatiotemporal sequence image registration tasks compared with current works.