TL;DR
This paper introduces OS-LALM, a novel algorithm for accelerating X-ray CT image reconstruction. By combining a linearized augmented Lagrangian method with ordered subsets and a deterministic downward continuation strategy, OS-LALM achieves significantly faster convergence and reduced artifacts compared to existing methods, even with a large number of subsets.
What Problem Does This Paper Address?
X-ray Computed Tomography (CT) image reconstruction is crucial for medical imaging, but traditional statistical methods, while offering better image quality and lower radiation doses, are computationally expensive. Existing methods like Augmented Lagrangian (AL) and Ordered Subsets (OS) have limitations. AL methods can be slow due to complex inner-problem solutions, especially in CT where the system matrix is "shift-variant." OS methods accelerate reconstruction but can introduce artifacts, particularly with many subsets. This paper aims to bridge this gap by developing a faster, more stable, and artifact-reducing CT reconstruction algorithm.
Methods
The core of the proposed method, OS-LALM, is built upon a linearized Augmented Lagrangian (AL) method. Instead of the standard AL approach, it uses a separable quadratic surrogate (SQS) function to approximate the quadratic penalty term. This simplification leads to a more manageable inner problem. Crucially, this linearized AL approach is made OS-accelerable, meaning it can leverage the speed-up offered by ordered subsets.
To further enhance convergence, the authors introduce a deterministic downward continuation approach. This strategy involves adaptively adjusting a key parameter (the AL penalty parameter, p) during the optimization process. By starting with a larger step size (smaller p) and gradually reducing it, the algorithm avoids the need for tedious manual parameter tuning and achieves faster convergence. The analysis of this approach is grounded in second-order recursive system analysis.
Datasets
The paper evaluates the proposed algorithm on two real-world CT datasets:
- A shoulder region helical CT scan.
- A GE performance phantom axial CT scan.
These datasets are used to compare the performance of OS-LALM against several state-of-the-art OS-based reconstruction algorithms.
Key Insights and Ideas
- Linearization is Key for OS Acceleration: The paper demonstrates that linearizing the AL penalty term is crucial for making the method compatible with ordered subsets, a key step towards practical acceleration in CT.
- Deterministic Downward Continuation: The proposed parameter tuning strategy avoids manual adjustments and provides a principled way to achieve fast convergence by adaptively controlling the step size. This is a significant advantage over methods requiring extensive parameter sweeps.
- Improved Artifact Reduction: OS-LALM shows a notable reduction in OS artifacts, especially when using a large number of subsets, which is a common challenge with standard OS methods. This suggests better gradient error tolerance.
- Single FISTA Iteration Suffices: For CT reconstruction, the inner constrained denoising problem can often be solved effectively with just one iteration of the FISTA algorithm, making the overall algorithm computationally efficient.
Key Results
- Significant Acceleration: OS-LALM achieves substantial acceleration in X-ray CT image reconstruction convergence. For example, in shoulder scan experiments, it showed approximate 5, 10, and 20 times acceleration compared to the standard OS algorithm with p = 0.2, 0.1, and 0.05, respectively.
- Reduced OS Artifacts: The algorithm significantly reduces OS artifacts, particularly when using many subsets (e.g., 40 subsets for the shoulder scan). Visual comparisons in Figure 2 and Figure 6 highlight this improvement.
- Robustness: The proposed algorithm performs well even with truncated data (truncated abdomen scan), exhibiting fast convergence and good gradient error tolerance.
- Efficiency: The method requires negligible overhead and can achieve excellent results with just one FISTA iteration for the inner problem.
Limitations and Future Work
- Convergence Analysis for M > 1: While the paper demonstrates convergence for the one-subset case and good performance for M > 1, a more rigorous convergence analysis for the OS-accelerated version with multiple subsets is suggested as future work.
- Parameter Tuning for Other Problems: The deterministic downward continuation approach is designed for CT reconstruction. Its applicability and effectiveness for other imaging modalities or inverse problems might require further investigation.
- High-Memory Version: For more challenging inner denoising problems, an extension involving an additional split variable for the regularizer might be necessary, leading to a "high-memory" version of OS-LALM.
Summary updated