Purpose: To better reconstruct 4D cone-beam computed tomography (CBCT) images, we develop a general simultaneous motion estimation and image reconstruction (G-SMEIR) method.
Methods: The original SMEIR updates the deformation vector fields (DVFs) using the projection-domain data fidelity and suffers a local optimum trapping problem. We proposed modified SMEIR to address this problem by re-initializing DVFs using image-domain motion estimation. In this work, we further develop G-SMEIR by using a flexible combination of SMEIR and Demons registration iterations. Specifically, SMEIR with the initial DVFs estimated from 3D reconstruction is applied to obtain images at each phase. Afterward, DVFs are updated using Demons registration of all phase images. Then, SMEIR with the updated DVFs is applied again. Two steps, image update using SMEIR and DVF update using Demons, repeats until it converges. To deal with the computationally intensive Demons registration, a GPU version is developed. The method is tested using a CBCT simulation study of 4D NCAT phantom.
Results: The proposed G-SMEIR method can effectively alleviate the local optimum trapping problem of SMEIR and reconstruct 4D images with better image quality than SMEIR and 3D reconstruction, even at half of the regular dose (â€œlow doseâ€?). G-SMEIR performed similarly at the regular and low doses. The root mean square error of G-SMIER is improved more than 60% over 3D reconstruction and 11%~17% over SMEIR. The structural similarity indices for a representative phase are 0.6418 (3D), 0.8893 (SMEIR), and 0.9206 (G-SMEIR). GPU Demons registration shortens computational time from 2 hours (CPU) to 5 minutes for seven pairs of 3D images.
Conclusion: We develop a G-SMEIR method for 4D CBCT with great potential to lower radiation dose. The repetition of computationally intensive Demons registration is addressed by GPU computing. The simulation study demonstrated that G-SMEIR achieves good image quality at half of the regular dose.
Funding Support, Disclosures, and Conflict of Interest: The work is supported in part by NIH R03 EB021600, R15 CA199020 and R01 EB020366.