Author: Weixing Cai, Laura I. Cervino, Yabo Fu, Bohong Huang, Licheng Kuo, Tianfang Li, Xiang Li, Jean M. Moran, Huiqiao Xie, Hao Zhang 👨🔬
Affiliation: Department of Medical Physics, Memorial Sloan Kettering Cancer Center, Memorial Sloan Kettering Cancer Center 🌍
Purpose: Four-dimensional cone-beam computed tomography (4D-CBCT) is critical in image-guided radiotherapy (IGRT) for visualizing tumor motion. However, sparse projection sampling often introduces severe streak artifacts and degrades image quality. This study aims to utilize the Denoising Diffusion Implicit Model (DDIM) to model the statistic distribution of a patient's anatomy, serving as an image prior for 1-minute 4D CBCT and improving reconstruction accuracy.
Methods: A patient specific DDIM was trained on 2D axial slices from the patient’s initial 4DCT dataset. A separate 4DCT, acquired on a different day, was used to simulate a 1-minute CBCT half-fan scan using Monte Carlo methods. Penalized Weighted Least Square (PWLS) reconstruction was applied to sparse projections within a phase bin to initialize the diffusion-based reconstruction. The diffusion model then generated a high-quality image constrained on the initial PWLS reconstruction, which served as a prior for subsequent PWLS iterations to enhance data consistency. Diffusion based reconstructions were compared against reconstructions using PWLS, and FDK visually, as well as quantitatively with Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).
Results: Diffusion-based reconstruction significantly outperformed both PWLS and FDK reconstructions. For 10 phases, each contains approximately 70 projections, diffusion-based reconstruction achieved a PSNR of 26.0 ± 0.12 and SSIM of 0.92 ± 0.001, compared to PWLS (PSNR: 25.4 ± 0.3, SSIM: 0.89 ± 0.004) and FDK (PSNR: 22.9 ± 1.6, SSIM: 0.87 ± 0.006). Our results demonstrated effective streaking artifacts suppression with improved structural fidelity. Reconstruction of a 256×256×110 CBCT volume with 2×2×1.25 mm voxels was completed in approximately 8 minutes using two Nvidia Quadro RTX 8000 GPUs.
Conclusion: Diffusion-based 4D CBCT reconstruction significantly improves in image quality over traditional FDK and PWLS methods. This innovative approach shows great potential for enabling accelerated 1-minute 4D CBCT imaging in radiotherapy, providing accurate visualization of tumor motion.