Geometric tomography: a limited-view approach for computed tomography

Peter B. Noël, Jinhui Xu, Kenneth R. Hoffmann, Jason J. CorsoSCG 2009 · DOI

TL;DR

This paper introduces a novel "geometric compressed sensing" approach to computed tomography (CT) that significantly reduces the number of X-ray projections needed for high-quality 3D reconstruction. By intelligently splitting the reconstruction problem into smaller, object-specific subproblems based on voxel intensity, the method aims to mitigate the radiation dose and computational burden associated with traditional CT.

What Problem Does This Paper Address?

Traditional CT imaging, while invaluable for medical diagnostics, faces two significant challenges: increasing patient radiation dose and lengthy reconstruction times. The need for high-quality 3D data often necessitates a large number of X-ray projections (hundreds to thousands), which directly correlates with higher radiation exposure and increased computational demands. While advancements like GPU acceleration can speed up reconstruction, the fundamental issue of radiation dose remains. Reducing the number of projections to lower dose can lead to artifacts and degraded image quality. This paper seeks to address this trade-off by developing a reconstruction method that can achieve good image quality with fewer projections.

Methods

The core of the proposed method is a "geometric compressed sensing" approach that leverages an observation about standard CT reconstruction: it converges quickly when voxel intensities are similar. The authors propose to mathematically split the general CT problem ($\tilde{y} = Mf$) into smaller subproblems, each focusing on an "Object-of-Interest" (OoI) with similar voxel intensity. This effectively creates object-specific reconstructions.

The approach is a two-stage process:

  1. OoI Isolation: The first stage automatically identifies OoIs (e.g., anatomical structures) from a limited-view reconstruction.
  2. Object-Specific Reconstruction: If OoIs are detected, the second stage uses these volumes to drive an independent reconstruction of each OoI and its complement. This involves classifying projection rays that penetrate through OoIs versus those that penetrate through their complements. To efficiently classify these rays, they employ a "point-line duality transformation" on boundary segments of the OoIs, creating a 2D arrangement that partitions space based on ray intersections. The "Topological Peeling" algorithm is then used to efficiently process these rays. Finally, these classified sparse ray sets are fed into a GPU-based reconstruction engine.

Datasets

Not discussed in the paper. The paper mentions using "rabbit data sets" for demonstrating reconstruction results, but no specific dataset names or details are provided.

Key Insights and Ideas

  • Exploiting Intensity Similarity: The key insight is that standard CT reconstruction algorithms perform well when voxel intensities are uniform. This observation is the foundation for splitting the problem into regions of similar intensity.
  • Object-Specific Reconstruction: By treating each region of similar intensity as a separate subproblem, the method allows for tailored reconstruction, potentially leading to better quality with fewer projections.
  • Geometric Compressed Sensing: The paper frames its approach as a form of geometric compressed sensing, where the geometric properties of the objects (defined by intensity regions) are used to guide the reconstruction.
  • Ray Classification via Duality: The use of point-line duality transformation to efficiently classify projection rays based on their interaction with object boundaries is a novel geometric approach to managing the complexity of ray-object interactions.

Key Results

The paper describes a "two stage technique" and a "geometric compressed-sensing approach" that aims to improve image quality with fewer projections. It demonstrates the reconstruction process using "rabbit data sets" and compares results with standard techniques. However, the paper does not provide specific quantitative results (e.g., percentage reduction in projections, improvement in image quality metrics like PSNR or SSIM) or concrete numbers for speedup. The qualitative finding is that their approach can achieve reconstruction from a limited number of projections.

Limitations and Future Work

  • Limitations: The paper does not explicitly state limitations, but the lack of quantitative results suggests that further validation and benchmarking are needed. The reliance on automatic OoI detection might be a point of failure if the segmentation is inaccurate.
  • Future Work: The paper implies future work through its description of the video sequence, which includes an animated explanation of the CT imaging setup, input acquisition, observations, the geometric compressed sensing approach, and reconstruction results. The mention of GPU-based reconstruction suggests ongoing work in leveraging hardware acceleration.

Summary updated

BibTeX
@inproceedings{noel2009geometric,
  author = {Peter B. Noël and Jinhui Xu and Kenneth R. Hoffmann and Jason J. Corso},
  title = {{Geometric tomography: a limited-view approach for computed tomography}},
  booktitle = {SCG},
  year = {2009},
  doi = {10.1145/1542362.1542383},
}
Figure