Cone beam computed tomography reconstruction from truncated projections using prior information and transfer learning

Cone beam computed tomography (CBCT) is widely used in clinical practice and small animal research for image guidance. The reconstruction quality will be compromised by severe truncation-related artifacts when the scanned object is not fully covered by the field of view (FOV).

This work aims to develop a Dual-Domain Deep learning-based method for CBCT Reconstruction from Truncated projections (D3CRT) through the guidance of non-truncated prior information.

The D3CRT comprised sequential procedures in both projection and image domains. First, in projection domain, a Sinogram Generation Network (SG-Net) based on the denoising diffusion probabilistic model (DDPM) was employed to predict the missing projection data outside the FOV. The SG-Net was fine-tuned via transfer learning using non-truncated prior data to achieve object-specific adaptation. FDK reconstruction was subsequently performed using the predicted projections. Second, in image domain, an Image Enhancement Network (IE-Net) was applied to refine the FDK reconstructed images. Compressed sensing (CS) reconstruction was then carried out to enforce data fidelity by incorporating the original projections, followed by a secondary IE-Net for final image quality enhancement. In-vivo small animal experiments were conducted on a micro-CBCT system to validate the D3CRT method, with non-truncated prior data obtained from large-FOV low-resolution scans. Dice similarity coefficient (DSC), structural similarity index measure (SSIM), root mean square error (RMSE) were used for quantitative evaluation.

The proposed D3CRT effectively improves the image reconstruction quality under truncated projection conditions. For whole-body and lung regions, D3CRT achieved DSCs of 97.1% and 96.0%, outperforming the low-resolution prior images (DSCs of 96.8% and 89.9%) when compared with the reference region segmentations. Quantitative evaluations within the FOV yielded an average RMSE of 2.95 $\times 10^{-3}m{m}^{-1}$ and an SSIM of 98.1% for D3CRT, demonstrating better performance than the Low Resolution Image Constrained Reconstruction (LRICR) method which directly takes low-resolution prior images as the initial inputs for CS reconstruction (RMSE 3.83 $\times {10}^{-3}m{m}^{-1}$ , SSIM 97.4%).

By leveraging non-truncated prior information and projection-domain transfer learning, the proposed D3CRT effectively improved the overall quality of CBCT reconstruction from truncated projections.