Deep Inertia Lp Half-Quadratic Splitting Unrolling Network for Sparse View CT Reconstruction

Yu Guo, Caiying Wu, Yaxin Li, Qiyu Jin, Tieyong Zeng

Introduction

Computed Tomography (CT) is an indispensable tool in clinical diagnostics. However, the ionizing radiation associated with standard CT scans presents significant health risks to patients. Sparse-view CT reconstruction aims to mitigate these risks by significantly reducing the number of X-ray projections. While beneficial for patient safety, this reduction leads to an "ill-posed" inverse problem, where traditional reconstruction methods often produce images marred by severe streak artifacts and high noise levels.

Recent advancements in deep learning have offered promising solutions for image denoising and artifact removal. Yet, many of these models function as "black boxes," lacking the theoretical convergence guarantees essential for medical applications. Conversely, classical iterative optimization methods are mathematically rigorous but computationally expensive and slow. This research introduces IHQS$_p$-Net, a deep unrolling network designed to bridge the gap between the speed of neural networks and the reliability of traditional optimization.

Methodology: The IHQS$_p$-Net Architecture

The core of the proposed method is the Inertial $L_p$-norm Half-Quadratic Splitting (IHQS$_p$) algorithm, which is "unrolled" into a deep neural network. This architecture allows the model to learn optimal parameters from data while adhering to a mathematically sound iterative structure.

$L_p$-norm Regularization and Inertial Acceleration

Unlike the standard $L_1$ norm, the $L_p$ norm (where $0 < p < 1$) provides a tighter approximation of true sparsity. This is crucial for recovering sharp edges and fine details in medical imaging. To address the slow convergence typical of iterative solvers, the authors integrate "inertial steps"—a concept borrowed from physics-inspired optimization (similar to momentum). These steps significantly accelerate the convergence process.

Smart CG Initialization

A major bottleneck in CT reconstruction is solving the "u-subproblem" using the Conjugate Gradient (CG) method. IHQS$_p$-Net optimizes this by using a Convolutional Neural Network (CNN), such as a UNet or DnCNN, to provide a high-quality initial guess for the CG solver. This "smart initialization" reduces the number of required iterations without compromising the theoretical convergence of the overall algorithm.

The IHQS$_p$-Net unrolling architecture, showing the iterative blocks for Proximal and CG updates with inertial acceleration.

Experimental Setup

The performance of IHQS$_p$-Net was rigorously evaluated using two primary datasets:

  1. 2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge: A standard clinical benchmark.
  2. Covid-19 Dataset: Used to assess the model's robustness across different lung pathologies.

The researchers simulated real-world scanner conditions by testing the models against both standard Gaussian noise and complex mixed noise (Gaussian + Poisson).

Key Results and Discussion

The proposed IHQS$_p$-Net consistently outperformed both traditional iterative methods and state-of-the-art deep learning models like PDNet and FBPNet.

Quantitative Performance

On the AAPM dataset with only 60 views, IHQS$_p$-Net achieved a Peak Signal-to-Noise Ratio (PSNR) of 30.24 dB. This is a massive improvement over the baseline Filtered Back Projection (FBP) at 18.21 dB. While competitive with models like MetaInvNet, IHQS$_p$-Net demonstrated superior structural similarity and fewer artifacts.

Qualitative Superiority

Visual inspections confirm that IHQS$_p$-Net preserves fine structural details that other models often over-smooth. Crucially, it avoids the "hallucinations" or pseudo-structures sometimes seen in purely data-driven models.

Visual comparison of reconstruction results across various methods. IHQS$_p$-Net (l) demonstrates superior detail preservation and artifact reduction compared to FBP (b) and other deep learning baselines.

Insights on Sparsity and Speed

Ablation studies revealed that $p=0.7$ is the "sweet spot" for balancing sparsity and optimization stability in CT images. Furthermore, the inclusion of inertial steps allowed the algorithm to reach optimal results in 65 iterations, compared to 90 iterations for the non-inertial version, representing a significant gain in efficiency.

Conclusion

IHQS$_p$-Net represents a significant step forward in sparse-view CT reconstruction by combining mathematical rigor with deep learning efficiency. By using a neural network as an initializer rather than a standalone solver, the authors ensure the output remains physically consistent with the CT measurement model. Future research will focus on automating hyperparameter selection and extending the framework to 3D cone-beam CT data.

Summary updated

BibTeX
@article{guo2024deep,
  author = {Yu Guo and Caiying Wu and Yaxin Li and Qiyu Jin and Tieyong Zeng},
  title = {{Deep Inertia Lp Half-Quadratic Splitting Unrolling Network for Sparse View CT Reconstruction}},
  journal = {IEEE Signal Processing Letters},
  year = {2024},
  doi = {10.1109/lsp.2024.3438118},
}
Figure