[Paper Review] Empirical average-case relation between undersampling and sparsity in x-ray CT

Jakob Sauer Jørgensen, Emil Y. Sidky+2 more|arXiv (Cornell University)|Nov 24, 2012

Medical Imaging Techniques and Applications35 references3 citations

TL;DR

This paper empirically establishes a quantitative, average-case relationship between image sparsity and the number of CT projections required for exact reconstruction via L1-minimization. Using simulated fan-beam CT data, it demonstrates a sharp phase transition in recoverability that depends on image class, is independent of image size, and remains robust under small Gaussian noise, providing practical guidance for low-dose CT imaging with sparsity-exploiting methods.

ABSTRACT

In x-ray computed tomography (CT) it is generally acknowledged that reconstruction methods exploiting image sparsity allow reconstruction from a significantly reduced number of projections. The use of such reconstruction methods is motivated by recent progress in compressed sensing (CS). However, the CS framework provides neither guarantees of accurate CT reconstruction, nor any relation between sparsity and a sufficient number of measurements for recovery, i.e., perfect reconstruction from noise-free data. We consider reconstruction through 1-norm minimization, as proposed in CS, from data obtained using a standard CT fan-beam sampling pattern. In empirical simulation studies we establish quantitatively a relation between the image sparsity and the sufficient number of measurements for recovery within image classes motivated by tomographic applications. We show empirically that the specific relation depends on the image class and in many cases exhibits a sharp phase transition as seen in CS, i.e. same-sparsity image require the same number of projections for recovery. Finally we demonstrate that the relation holds independently of image size and is robust to small amounts of additive Gaussian noise.

Motivation & Objective

• To empirically determine the relationship between image sparsity and the number of CT projections sufficient for exact reconstruction using L1-minimization.
• To investigate whether this relationship depends on image class, size, or noise level.
• To provide quantitative, data-driven guidance for low-dose CT reconstruction with sparsity-exploiting methods.
• To explore the robustness of the sparsity-sampling relation under realistic conditions such as additive noise and varying image structure.

Proposed method

• Simulated fan-beam CT data were generated using a standard CT geometry with a fixed number of projections.
• Sparse images were synthesized from well-defined classes (e.g., random, 2-power, blocky) with controlled sparsity levels.
• Reconstruction was performed via L1-minimization subject to data consistency constraints, solved using the MOSEK optimization solver for robustness.
• Recovery was assessed by comparing reconstructed images to original sparse images using a strict error threshold.
• Phase diagrams were constructed to visualize the transition from non-recovery to recovery as a function of sparsity and sampling rate.
• Robustness was tested by introducing small additive Gaussian noise into the projection data.

Experimental results

Research questions

• RQ1 What is the quantitative relationship between image sparsity and the number of CT projections required for exact reconstruction via L1-minimization?
• RQ2 How does this relationship vary across different image classes with distinct structural characteristics?
• RQ3 Is the sparsity-sampling relation independent of image size, and does it hold for larger images?
• RQ4 How robust is the recovery relation to small amounts of additive Gaussian noise in the projection data?
• RQ5 Does the phase transition from non-recovery to recovery exhibit sharpness or gradualness, and what factors influence this?

Key findings

• A clear, average-case relation exists between image sparsity and the number of projections required for exact L1-reconstruction in fan-beam CT.
• The recovery transition is sharp for less-structured image classes, such as those with random or 2-power sparsity patterns.
• For more structured image classes (e.g., blocky), the required number of projections is smaller but exhibits higher variability in recovery.
• The sparsity-sampling relation is independent of image size, suggesting scalability through extrapolation.
• The relation remains robust under small amounts of additive Gaussian noise, indicating practical relevance for real-world CT systems.
• The findings suggest the existence of an underlying theoretical foundation for compressed sensing in CT that remains to be formally established.