Purpose: Four-dimensional computed tomography (4DCT) provides important physiological information for diagnosis and treatment. On the other hand, its acquisition is challenged with artifacts due to motion or sorting/binning, time and effort bandwidth, and dose considerations. A 4D synthesis development would significantly augment the available data, addressing quality and consistency issues. Furthermore, a high-quality synthesis can serve as an essential backbone to establish a feasible physiological manifold to support online reconstruction, registration, and downstream analysis from real-time x-ray imaging. In this study, we propose a novel approach to synthesize continuous 4D respiratory motion from two extreme respiration phases.
Methods: A conditional image registration network is trained to take two breathing phases as input, and output arbitrary breathing phases. The network has a general U-net architecture and takes an additional variable input to the encoding path to provide phase conditioning for the network. The network is trained on 4DCT scans in an unsupervised fashion and can generate 4D continuous motion fields by varying the conditional variable.
Results: The method was tested on 20 4DCT scans and demonstrated to generate realistic 4D respiratory motion fields that were spatiotemporally smooth, achieving a root-mean-square error of (72.6 ± 34.1) HU and structural similarity index of (0.925 ± 0.046), compared to the ground-truth 4DCT. A 10-phase synthesis takes about 2.85 s.
Conclusion: We demonstrated that the spatiotemporal non-linearity of respiratory motion can be modeled by a conditional registration network. The proposed calibration module can correct for regularization-induced phase bias effectively. Our method can synthesize realistic 4D continuous motion fields and can be applied in common clinical practice to support 4DCT synthesis, online reconstruction, and other downstream tasks.