3D diffusion posterior sampling for CT reconstruction

Diffusion models have demonstrated a powerful capability to generate a diversity of high quality images based on a training distribution. Recently, such diffusion models have been used in CT restoration and reconstruction via conditional generation. Diffusion posterior sampling (DPS) is a conditional generation method with several advantages, including unsupervised learning of the prior distribution and plug-and-play capabilities with different forward models to encompass different acquisition methods, protocols, etc. However, most current DPS work has focused on two-dimensional models for both the prior and system models. Almost all clinical CT systems are inherently three-dimensional using helical or cone-beam acquisitions. While the extension to 3D is mathematically straightforward, computational demands prohibit direct application on most platforms. In this research, we propose strategies for 3D DPS CT reconstruction using a 3D neural network to learn the prior distribution. We develop modifications to a standard DPS algorithm to substantially reduce memory requirements and to accelerate the sampling speed. We evaluate different alternatives that permit 3D DPS in realistic CT volume sizes and compare relative merits of each strategy.