Neighbor-based adaptive sparsity orthogonal least square for fluorescence molecular tomography

Huangjian Yi, Sihao Ma, Ruigang Yang, Sheng Zhong, Hongbo Guo, Xuelei He, Xiaowei He, Yuqing HouJournal of Biomedical Optics 2023 · DOI

TL;DR

This paper introduces NASOLS, a novel algorithm for Fluorescence Molecular Tomography (FMT) that improves image reconstruction quality by adaptively determining sparsity and leveraging neighbor information from the mesh. This advancement is crucial for better visualizing biological processes in vivo, with potential applications in early tumor detection.

What Problem Does This Paper Address?

Fluorescence Molecular Tomography (FMT) is a powerful imaging technique for visualizing biological processes at the molecular level in living organisms. However, FMT reconstruction is notoriously challenging due to the strong scattering of light within tissues and the fact that the number of measurements is often far smaller than the number of unknowns. This makes the problem "ill-posed," meaning small errors in measurements can lead to large errors in the reconstructed image, often resulting in blurry or inaccurate representations of fluorescent targets. Existing methods often struggle with accurately localizing and resolving multiple targets, especially when the sparsity (i.e., the number of active fluorescent sources) is unknown beforehand.

Methods

The core innovation of this paper is the Neighbor-based Adaptive Sparsity Orthogonal Least Squares (NASOLS) algorithm. Here's a breakdown of its key components:

  • Adaptive Sparsity: Unlike traditional methods that require prior knowledge of the sparsity level, NASOLS adaptively determines it. It uses a nonlinear function to adjust the step size for increasing sparsity, allowing it to quickly approach the true sparsity in the initial stages and then refine it more gradually.
  • Neighbor Strategy: To preserve spatial information and improve reconstruction, NASOLS incorporates a "neighbor strategy." This strategy leverages the finite element mesh structure. For each identified fluorescent source (represented as a node in the mesh), it considers its neighboring nodes (connected by tetrahedral edges) when expanding the support set (the set of potential locations for fluorescent sources). This helps maintain the local structure of the fluorescence distribution.
  • Orthogonal Least Squares (OLS): NASOLS builds upon the OLS algorithm, which iteratively selects the most relevant components to reconstruct the signal. NASOLS enhances OLS by integrating its adaptive sparsity and neighbor strategy for more robust support set selection.
  • L1-norm Regularization: The underlying mathematical model uses L1-norm regularization, a common technique in compressed sensing that promotes sparse solutions, aligning well with the assumption that fluorescent targets are often localized and sparse within the larger imaging volume.

Datasets

The paper evaluates NASOLS using:

  • Numerical Simulations: A 3D digital mouse model was used to simulate FMT data.
  • Physical Phantom Experiments: A cubic phantom with two precisely placed cylindrical targets was used.
  • Small Animal In-vivo Experiments: Experiments were conducted on two small mice with implanted fluorescent targets.

The paper does not introduce new benchmark datasets but rather uses these established experimental setups and models to compare NASOLS against existing methods.

Key Insights and Ideas

  • Sparsity is Key, but Unknown: The paper highlights that while FMT targets are often sparse, knowing the exact sparsity beforehand is a major hurdle. NASOLS's adaptive approach directly addresses this.
  • Local Structure Matters: The "neighbor strategy" is a crucial insight. By considering the spatial relationships between potential fluorescent sources in the mesh, NASOLS can better preserve the integrity of the reconstructed targets, especially when multiple targets are close together.
  • Double-Target Resolution: The method shows a significant improvement in distinguishing closely spaced fluorescent targets, a common challenge in FMT.
  • Robustness to Noise and Projections: NASOLS demonstrates good performance across varying numbers of projection measurements and levels of noise, suggesting its practical applicability.

Key Results

NASOLS consistently outperformed other methods (ASOLS, OLS, CoSaMP, gOMP) across various experiments:

  • Single Target Reconstruction: NASOLS achieved the smallest Location Error (LE) and Normalized Root-Mean-Square Error (NRMSE), and the largest Contrast-to-Noise Ratio (CNR) in numerical simulations. For example, in one simulation, NASOLS achieved an LE of 0.41 mm, NRMSE of 0.0012, and CNR of 2.29, outperforming other methods.
  • Double Target Reconstruction: NASOLS was able to distinguish two targets even when they were closely spaced (EED of 2 mm), a feat that other methods struggled with or failed to achieve. For an EED of 4 mm, NASOLS achieved the smallest LE, NRMSE, and largest CNR.
  • In-vivo Experiments: In small animal studies, NASOLS also showed superior performance in accurately locating targets and providing better overall image quality compared to the other algorithms. For Mouse 1, NASOLS achieved an LE of 1.2 mm and CNR of 0.49, which were the best among the tested methods.

Limitations and Future Work

The authors acknowledge a few limitations and suggest future directions:

  • Neighbor Node Repetition: The selection of nearest neighbor nodes can sometimes be repeated, potentially increasing computation time.
  • Shape Reconstruction: While NASOLS improves target localization, its ability to reconstruct the precise shape of the target is still limited, which is a significant challenge in FMT.
  • Future Work: The authors plan to continue working on improving the shape reconstruction capabilities of FMT. They also note that while deep learning methods are emerging, traditional methods like NASOLS still offer advantages, particularly in their interpretability and ability to incorporate physical models.

Summary updated

Abstract

NASOLS can recover the fluorescence target with a good location error according to simulation experiments, phantom experiments and small mice experiments. This method is suitable for sparsity target reconstruction, and it would be applied to early detection of tumors.
BibTeX
@article{yi2023neighbor,
  author = {Huangjian Yi and Sihao Ma and Ruigang Yang and Sheng Zhong and Hongbo Guo and Xuelei He and Xiaowei He and Yuqing Hou and Xiaowei He and Yuqing Hou},
  title = {{Neighbor-based adaptive sparsity orthogonal least square for fluorescence molecular tomography}},
  journal = {Journal of Biomedical Optics},
  year = {2023},
  doi = {10.1117/1.jbo.28.6.066005},
}
Figure