Mixed one-bit compressive sensing with applications to overexposure correction for CT reconstruction

Xiaolin Huang, Yan Xia, Lei Shi, Yixing Huang, Ming Yan, Joachim Hornegger, Andreas K. MaierarXiv 2017 · arXiv

Introduction

In medical imaging, particularly with C-arm Computed Tomography (CT) systems, flat-panel detectors often struggle with a limited dynamic range. When X-rays pass through thin anatomical regions or strike the detector directly, the resulting signal can exceed the sensor's maximum measurable value, leading to saturation or overexposure. Traditionally, these saturated pixels are treated as lost data, resulting in significant reconstruction artifacts. This paper introduces Mixed One-bit Compressive Sensing (M1bit-CS), a novel framework that reinterprets saturated measurements as valuable inequality constraints rather than discarded noise.

The Challenge of Saturation

Standard reconstruction algorithms, such as Filtered Back Projection (FBP), are highly sensitive to missing or clipped data. When saturation occurs, the reconstructed images typically suffer from three primary issues:

  1. Streaking artifacts: Sharp, radiating lines originating from dense objects.
  2. Capping artifacts: A phenomenon where homogeneous regions appear darker at the edges than in the center.
  3. Loss of boundaries: The outer edges of the patient's anatomy may disappear entirely.
Figure 2: (a) Original phantom, (b) FBP reconstruction showing artifacts, (c) original sinogram, and (d) saturated sinogram illustrating the data loss.

While hardware solutions like bowtie filters or mathematical extrapolation techniques exist, they often fail to fully recover the lost structural information. The M1bit-CS approach instead leverages the fact of saturation: a saturated pixel indicates that the true signal is at least as large as the detector's threshold.

Proposed Method: M1bit-CS

The core innovation of M1bit-CS is the combination of standard analog measurements with "one-bit" information derived from saturated pixels. The reconstruction is formulated as a convex optimization problem using two primary loss functions:

  • Least-squares loss: Applied to regular, unsaturated measurements.
  • Pinball loss: A generalization of hinge loss used for saturated measurements. This function penalizes reconstructed values that fall below the saturation threshold while remaining robust to noise.

The authors propose two variants: M1bit-CSC, which utilizes an $\ell_2$-norm constraint (ideal when signal energy is known), and M1bit-CSR, which uses $\ell_2$-norm regularization for more general clinical applications. To solve these problems efficiently, the researchers developed an algorithm based on the Alternating Direction Method of Multipliers (ADMM). Additionally, an Iterative Saturation Detection (ISD) scheme was introduced to distinguish between true saturation and instances where X-rays hit the detector directly without passing through tissue.

Experimental Validation

The framework was evaluated across three distinct scenarios:

  • 1D Synthetic Signals: To benchmark Signal-to-Noise Ratio (SNR) recovery performance against traditional compressive sensing methods.
  • Simulated Knee Phantom: To visualize the mitigation of capping and streaking artifacts in a controlled environment.
  • Clinical Head Dataset: Using real data from a Siemens Artis zee angiographic C-arm system to simulate heavy overexposure scenarios.

Results and Discussion

The quantitative and qualitative results demonstrate a significant leap over existing methods. In the clinical head dataset with heavy saturation, M1bit-CS achieved a Root Mean Square Error (RMSE) of 14.81 HU, a dramatic improvement over standard FBP (132.2 HU) and water-cylinder extrapolation (87.93 HU). Qualitative results on phantoms showed that M1bit-CS effectively eliminated the "capping" effect and restored outer boundaries.

Figure 6: (a) SNR performance comparison and (b) computation time vs. number of measurements for various algorithms.

As shown in Figure 6, both M1bit-CS variants consistently outperform standard Lasso and Robust Dual Compressive Sensing (RDCS) in terms of SNR as the number of measurements increases. However, this accuracy comes at a computational cost; the ADMM-based iterative approach is significantly slower than analytical FBP, as evidenced by the time-to-measurement plots.

Conclusion

Mixed One-bit Compressive Sensing represents a paradigm shift in handling overexposed CT data. By treating saturation as a mathematical bound rather than a failure of the sensor, the M1bit-CS framework successfully restores anatomical boundaries and eliminates pervasive artifacts. Future work will focus on reducing computational complexity and incorporating anatomical priors to move this technology closer to real-time clinical implementation.

Summary updated

BibTeX
@misc{huang2017mixed,
  author = {Xiaolin Huang and Yan Xia and Lei Shi and Yixing Huang and Ming Yan and Joachim Hornegger and Andreas K. Maier},
  title = {{Mixed one-bit compressive sensing with applications to overexposure correction for CT reconstruction}},
  year = {2017},
  eprint = {1701.00694},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1701.00694},
}
Figure