Fusing Sparsity with Deep Learning for Rotating Scatter Mask Gamma Imaging

Yilun Zhu, Clayton Scott, Darren E. Holland, George V. Landon, Aaron P. Fjeldsted, Azaree T. Lintereur2022 · DOI

TL;DR

This paper introduces a novel image reconstruction technique for Rotating Scatter Mask (RSM) gamma imaging systems, crucial for nuclear safety applications. By combining traditional sparsity-promoting methods with a deep learning-based denoiser, the authors achieve significantly improved image reconstruction accuracy compared to existing approaches, addressing the challenge of limited measurement data.

What Problem Does This Paper Address?

Accurately locating radioactive materials is vital for nuclear safety, whether for environmental monitoring or searching for orphan sources. Portable imaging systems are increasingly needed for this task. The Rotating Scatter Mask (RSM) system is a promising technology for this purpose, offering a near 4π field-of-view. However, a major hurdle in RSM imaging is the "under-determined" nature of the data acquisition. This means the system captures far fewer measurements (signals) than there are pixels in the image it aims to reconstruct. Effectively, it's like trying to paint a detailed picture using only a handful of broad strokes. Traditional methods struggle with this severe data limitation, leading to inaccurate or artifact-laden images.

Methods

The core of this work lies in a novel image reconstruction algorithm that fuses two powerful approaches:

  • Model-based Sparsity-Promoting Regularization: This leverages the inherent property that radiation sources are often sparse (i.e., concentrated in small areas). The paper uses an $l_1$ penalty, a standard technique in compressed sensing, to encourage sparse solutions.
  • Data-driven Deep Neural Network (CNN) Denoising Prior: Instead of relying solely on analytical models, the authors incorporate a pre-trained Convolutional Neural Network (CNN) as an "image prior." This CNN acts as a sophisticated denoiser, learning to produce cleaner and more realistic image structures from noisy data.

These two priors are integrated within an Alternating Direction Method of Multipliers (ADMM) framework. ADMM is a powerful optimization technique that breaks down a complex problem into smaller, more manageable sub-problems. The authors specifically employ a "Plug-and-Play" ADMM variant, which allows for the seamless integration of advanced, pre-trained denoisers (like their CNN) into the optimization process. This approach is novel because it applies deep denoiser priors to sparse images, a scenario not previously addressed by such methods. The algorithm is further optimized for computational efficiency using Fourier transforms, reducing the complexity of matrix inversions.

Datasets

Not discussed in the paper. The authors used synthetically generated noisy data for evaluation.

Key Insights and Ideas

  • Synergy of Sparsity and Deep Learning: The paper demonstrates that combining traditional sparsity priors with modern deep learning denoisers yields superior results. Sparsity helps in identifying the locations of sources, while the CNN denoiser refines the image structure and reduces noise, leading to a more accurate overall reconstruction.
  • Plug-and-Play ADMM for RSM: The ADMM framework, particularly the Plug-and-Play variant, proves to be a flexible and effective tool for integrating diverse regularization techniques (analytical sparsity and data-driven CNN) into a unified reconstruction algorithm.
  • Addressing the Under-determined Problem: The proposed method directly tackles the fundamental challenge of limited measurements in RSM imaging by effectively leveraging prior knowledge about the image content.

Key Results

The proposed $l_1$-CNN method significantly outperforms existing approaches, including Maximum Likelihood Expectation-Maximization with Median Root Prior (MLEM-MRP) and a pure $l_1$ reconstruction method.

  • Quantitative Improvement: On average, the proposed method improved the Normalized Root Mean Square Error (NRMSE) by over 30% compared to baseline methods. For example, the average NRMSE was 0.553 for $l_1$-CNN, compared to 1.31 for $l_1$ and 0.794 for MLEM-MRP.
  • Qualitative Improvement: Visual inspection of reconstructed images shows that the $l_1$-CNN method produces clearer and more artifact-free images, especially for complex shapes like squares.

Limitations and Future Work

  • Artifacts in Square Sources: While NRMSE was low for square sources, visual artifacts were still present, suggesting room for further refinement.
  • Future Directions: The authors suggest exploring "deep unrolling" techniques, which involve training the entire iterative algorithm architecture end-to-end, as a promising avenue for further improvement. This could lead to even more specialized and effective reconstruction models.

Summary updated

Abstract

Many nuclear safety applications need fast, portable, and accurate imagers to better locate radiation sources. The Rotating Scatter Mask (RSM) system is an emerging device with the potential to meet these needs. The main challenge is the under-determined nature of the data acquisition process: the dimension of the measured signal is far less than the dimension of the image to be reconstructed. To address this challenge, this work aims to fuse model-based sparsity-promoting regularization and a data-driven deep neural network denoising image prior to perform image reconstruction. An efficient algorithm is developed and produces superior reconstructions relative to current approaches.
BibTeX
@misc{zhu2022fusing,
  author = {Yilun Zhu and Clayton Scott and Darren E. Holland and George V. Landon and Aaron P. Fjeldsted and Azaree T. Lintereur},
  title = {{Fusing Sparsity with Deep Learning for Rotating Scatter Mask Gamma Imaging}},
  year = {2022},
  doi = {10.1109/nss/mic44845.2022.10399334},
}
Figure