3D Diffusion Posterior Sampling for CT Reconstruction.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Diffusion models excel at high-quality image generation.
• Diffusion Posterior Sampling (DPS) offers unsupervised learning and flexibility for CT restoration and reconstruction.
• Current DPS methods are primarily limited to 2D, while clinical CT is inherently 3D.

Purpose of the Study:

• To develop and evaluate strategies for 3D Diffusion Posterior Sampling (DPS) CT reconstruction.
• To address the computational challenges of applying 3D DPS to realistic CT volumes.
• To enable efficient and effective 3D CT reconstruction using advanced generative models.

Main Methods:

• Implementation of a 3D neural network for learning the prior distribution in DPS.
• Modification of the standard DPS algorithm to reduce memory usage and increase sampling speed.
• Evaluation of alternative strategies for enabling 3D DPS on realistic CT volume sizes.

Main Results:

• Development of computationally feasible strategies for 3D DPS CT reconstruction.
• Significant reduction in memory requirements and acceleration of sampling speed for 3D DPS.
• Comparative analysis of different strategies for 3D DPS implementation in CT.

Conclusions:

• The proposed strategies enable practical 3D DPS for CT reconstruction, overcoming previous computational barriers.
• The research facilitates the application of advanced generative models to complex 3D medical imaging tasks.
• This work paves the way for more efficient and accurate 3D CT reconstruction using diffusion models.