1150 - A Zero-Shot Deep Learning Method for 2V-CBCT with a Plug-and-Play Diffusion Prior

Purpose/Objective(s): Cone-beam computed tomography (CBCT) is extensively utilized in radiation therapy (RT) for patient setup. However, full-view projection data are not acquired for all RT fractions, whereas two orthogonal projections are consistently available for patient positioning. In this context, two-view CBCT (2V-CBCT) offers a promising approach for reconstructing volumetric imaging to enable dose reconstruction and accumulation. The reconstruction problem is a highly challenging task without the help of anatomical prior to address the under-determination. While supervised learning has shown feasibility in 2V-CBCT reconstruction using paired datasets, such methods face generalization challenges and depend on well-aligned paired data, which is often unavailable in practice. We propose a zero-shot method leveraging pretrained diffusion models for 2V-CBCT reconstruction without requiring model fine-tuning. The diffusion models serve as a strong prior, compensating for the limited information in 2V-CBCT projections for faithful reconstruction.

Materials/Methods: The diffusion models were pretrained on filtered back-projection (FDK) reconstructions of 700 pelvic patients using 720 equispaced projections. To improve training efficiency, we separately trained two 2D diffusion models using orthogonal 2D sliced images derived from the 3D volumetric image, instead of directly training a 3D diffusion model. These pretrained 2D models were then integrated into a traditional regularized optimization framework in a plug-and-play (PnP) manner. The optimization objective consisted of a data fidelity term to ensure consistency with the measured projections and a PnP diffusion term to enforce anatomical priors. A forward-backward splitting method was employed, allowing the data fidelity and diffusion prior terms to be addressed separately. The pretrained diffusion models acted as denoisers to guide intermediate reconstructions from measurement-guided gradient descent toward the anatomical manifold.

Results: The proposed method was evaluated on an independent dataset of 15 patients with the same pelvic site using metrics, including peak signal-to-noise ratio (PSNR) and multiscale structural similarity index measure (MS-SSIM). Compared to traditional FDK and a supervised 3D U-Net approach, our method demonstrated superior performance. For axial slices, the method achieved an average PSNR of 28.57 and MS-SSIM of 0.93, significantly outperforming the 3D U-Net (PSNR: 24.83, MS-SSIM: 0.87).

Conclusion: The proposed zero-shot deep learning method enables accurate, high-quality 2V-CBCT reconstruction from only two projection views. By leveraging the powerful plug-and-play diffusion prior, this approach holds significant potential for reliable application in radiotherapy.