Author: Xiaoyi Dai, Manju Liu, Weiwei Sang, Pulin Sun, Fan Xia, Zhenyu Yang, Fang-Fang Yin, Chulong Zhang, Rihui Zhang 👨🔬
Affiliation: Jiahui International Hospital, Radiation Oncology, Duke Kunshan University, Medical Physics Graduate Program, Duke Kunshan University 🌍
Purpose:
Current deep learning-based sparse-view CBCT reconstruction methods are prone to hallucinatory artifacts, as they rely on inferred image details that may not correspond to true anatomical projections. This study proposes a novel dual-branch network architecture that integrates planning CT as prior information to improve quality and fidelity of sparse-view CBCT reconstructions.
Methods:
We hypothesized that the high-resolution structural information in planning CT images could be integrated into sparse-view CBCT reconstruction to improve image quality while reducing the number of required projections. To test this hypothesis, we developed a dual-branch 3D UNet CNN that process both sparse-view CBCT projections and planning CT image patches. The self-attention and cross attention modules were innovatively employed to (1) enhance each branch’s feature characterization and (2) exchange information between branches, respectively. The final CBCT image was obtained by fusing feature information from both branches.
The study utilized a dataset of 637 CBCT scans from 32 patients, and each patient includes planning CT and multiple CBCT scans (~210 projections per scan). Unlike previous studies that primarily rely on simulated projections, our dataset was based on real clinically acquired CBCT projections. Sparse-view CBCT reconstructions were generated using 1/8 of the original projections. A 4:1 train-test split was applied. Reconstruction performance was evaluated against full-view reconstructions using PSNR and SSIM. Comparative analyses were conducted against classic 2D and 3D U-Net that directly process sparse-view CBCT projections.
Results:
The proposed model achieved superior performance (PSNR=35.92dB/SSIM=0.88), significantly outperforming both 2D(PSNR=34.67dB/SSIM=0.81) and 3D (PSNR=34.89dB/SSIM=0.83) U-Net models. Visual assessments further confirmed a significant reduction in artifacts and improved retention of subtle anatomical features compared to other methods.
Conclusion:
This study demonstrates the effectiveness of integrating planning CT as prior knowledge to enhance CBCT reconstruction. The proposed dual-branch network with self- and cross-attention offers an efficient solution for achieving high-quality CBCT imaging.