Vvbpnet: Deep Learning Model in View-by-View Backprojection (Vvbp) Domain for Sparse-View Cbct Reconstruction

Objective: To improve the quality of sparse-view cone-beam computed tomography (CBCT) images, a deep learning model in the view-by-view backprojection (VVBP) domain, VVBPNet, is proposed.Methods: The VVBPNet model adopted a content-noise complementary learning strategy, featuring two parallel attention Res-UNet sub-networks and a fusion mechanism. It processed VVBP-Tensors, which were intermediate results generated during the execution of the Feldkamp-Davis-Kress algorithm, to obtain denoised and artifact-reduced axial images. The model was trained, validated, and tested on CBCT data from 163, 30, and 30 real patients, respectively. Quantitative metrics including root-mean-square error (RMSE), peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and feature similarity (FSIM) were calculated to evaluate the performance of the model. The VVBPNet model was compared with 14 state-of-the-art (SOTA) models on three reconstruction tasks with varying sparsity levels: 1/4 moderate-sparse-view, 1/8 high-sparse-view, and 1/16 ultra-sparse-view.Results: For 1/4 task, the averaged metrics obtained by the VVBPNet model showed differences of -0.00001 RMSE, +0.1 dB PSNR, -0.002 SSIM, and –0.003 FSIM compared to the best averaged metrics from all SOTA models. For 1/8 task, VVBPNet showed differences of -0.00003 RMSE, +0.3 dB PSNR, +0.005 SSIM, and +0.001 FSIM compared to all SOTA models. For 1/16 , VVBPNet showed differences of -0.00009 RMSE, +0.6 dB PSNR, +0.009 SSIM, and +0.008 FSIM compared to all SOTA models.Conclusion: The proposed VVBPNet model in the VVBP domain effectively improves the quality of sparse-view CBCT images. As the projection views become sparser, the VVBPNet model exhibits greater performance advantages over the SOTA models.