Hybrid deep learning reconstruction for fast four-dimensional cone beam computed tomography in small animal imaging

Conventional four-dimensional cone beam computed tomography (4D-CBCT) usually requires long scan time and high radiation dose. Fast and low-dose 4D-CBCT is preferred but its quality is compromised by the reduced number of projections used for reconstruction.

This study is aimed to develop a hybrid deep learning reconstruction method for fast 4D-CBCT (HDR-4D) which can improve image quality while reducing imaging dose and scanning time.

The HDR-4D was initiated with the FDK reconstruction of each phase, followed by deep learning MKB (DL-MKB) method to preliminarily remove the streak artifacts in 4D-CBCT images. Then motion compensation (MoCo) was applied to fuse DL-MKB images of all phases. Finally, data fidelity was enforced under the framework of prior image constrained compressed sensing (PICCS) with the MoCo outputs as the prior information. An adaptive bone weighting strategy was introduced into the PICCS procedure to mitigate the bone-induced streak artifacts. A neural network SARnet was developed to remove the streak artifacts in the HDR-4D workflow. In-vivo animal imaging was conducted to validate the proposed method. Root mean square error (RMSE) and structure similarity index measure (SSIM) were calculated for quantitative evaluation.

The proposed method could reduce the 4D-CBCT scan time to 45 s. The reconstructed images exhibit no discernible streak artifacts. Compared to direct deep learning prediction, the proposed method demonstrated significantly enhanced performance in preserving detailed anatomical structures. The average RMSE was 2.87 $\times {10}^{-3}m{m}^{-1}$ and SSIM was 0.982 for 4D-CBCT images reconstructed with the HDR-4D method, outperforming the direct deep learning prediction (RMSE: 3.96 $\times {10}^{-3}m{m}^{-1}$ , SSIM: 0.966).

The proposed HDR-4D can generate high quality 4D-CBCT images while significantly reducing the scan dose and time, and demonstrates strong potential for small animal imaging applications.