Deep Learning Filter Replacement for Sparse-View CBCT Reconstruction: A Comparative Study of Image-Domain Residual U-Net Enhancement and Projection-Domain Learnable Filtering within Differentiable Feldkamp–Davis–Kress Algorithm
Abstract
Purpose
Conventional filtered backprojection with a fixed Ram-Lak filter in cone-beam CT (CBCT) reconstruction method often amplifies noise and streak artifacts under sparse-view acquisition, limiting image quality for image-guided radiotherapy. This study investigates deep learning–based filter replacement strategies and compares two complementary approaches: an image-domain network that replaces Ram-Lak filter and a projection-domain learnable filter embedded within a differentiable FDK framework.
Methods
High-quality diagnostic CT volumes were forward-projected to generate sparse-view CBCT projection data and served as ground truth for evaluation. For the image-domain approach, projections were first reconstructed without filtering to obtain low-quality Raw-CBCT images. A supervised ResUNet convolutional neural network was trained to enhance Raw-CBCT images as the predicting Ground Truth-CT. For the projection-domain approach, projection data were first processed by a UNet-based filtering network trained to learn data-adaptive projection-domain weighting in place of the conventional Ram-Lak filter. The learned projection-domain filter was then incorporated into a differentiable FDK reconstruction pipeline, enabling end-to-end optimization of filtering and backprojection to generate reconstructed CBCT volumes. Performance was quantitatively evaluated by comparing conventional FDK reconstructions and deep learning–based reconstructions against Ground-Truth CT using mean squared error (MSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR).
Results
The reconstructed CBCT images obtained using the proposed deep learning–based reconstruction frameworks demonstrated substantial improvements over conventional FDK. For two test cases, MSE was reduced by 80.0% and 63.4%, SSIM increased by 5.3% and 4.3%, and PSNR improved by 17.3 dB and 10.6 dB, respectively. Visual assessment showed effective streak suppression, noise reduction, and improved soft-tissue contrast.
Conclusion
The proposed image-domain and projection-domain deep learning approaches successfully replace conventional Ram-Lak filtering in CBCT reconstruction, achieving improved image quality compared with conventional FDK reconstruction.