3D CT Reconstruction from Few X-rays

Background

Computed Tomography (CT) imaging faces a persistent challenge in balancing the need for high-quality, accurate three-dimensional reconstructions with the imperative to minimize patient radiation exposure and scanning time. Traditional reconstruction algorithms, such as filtered back projection, are computationally efficient but are prone to generating noise and artifacts, particularly when fewer X-rays are used to reduce dose. While more advanced iterative reconstruction methods can mitigate some of these image quality issues, they often come with increased computational demands and longer processing times. The fundamental difficulty lies in accurately synthesizing detailed 3D volumetric data from a limited number of 2D X-ray projections, which inherently introduces ambiguity and uncertainty, potentially leading to blurry or less precise reconstructions.

Technology

Researchers at Stony Brook University developed a technology that reconstructs CT images by integrating implicit neural representation and conditional diffusion models. It begins with a nongenerative image-conditioned CT reconstruction using perceptual loss, then employs conditional diffusion models to address blurriness arising from limited input images. A neural voxel feature field is introduced for CT volume representation, which is formed by backprojecting local image features onto corresponding voxels and aggregating 2D features from multiple images. A conditional diffusion model, integrated with this 3D geometry-aware neural feature field, then samples the CT volume to synthesize high-fidelity and geometrically consistent 3D volumes.