Sparse Signal Reconstruction for Overdispersed Low-Photon Count Biomedical Imaging Using ℓp Total Variation

Yu Lu, Roummel F. MarciaarXiv 2024 · DOI · arXiv

Abstract

The negative binomial model, which generalizes the Poisson distribution model, can be found in applications involving low-photon signal recovery, including medical imaging. Recent studies have explored several regularization terms for the negative binomial model, such as the ℓp quasi-norm with 0 < p < 1, ℓ1 norm, and the total variation (TV) quasi-seminorm for promoting sparsity in signal recovery. These penalty terms have been shown to improve image reconstruction outcomes. In this paper, we investigate the ℓp quasi-seminorm, both isotropic and anisotropic ℓp TV quasi-seminorms, within the framework of the negative binomial statistical model. This problem can be formulated as an optimization problem, which we solve using a gradient-based approach. We present comparisons between the negative binomial and Poisson statistical models using the ℓp TV quasi-seminorm as well as common penalty terms. Our experimental results highlight the efficacy of the proposed method.

Introduction

In advanced imaging applications like low-dose medical CT scans or deep-space astronomy, the number of photons reaching the detector is often extremely low. Traditionally, the Poisson distribution is used to model this "shot noise." However, the Poisson model assumes that the variance equals the mean. In practice, many imaging systems exhibit overdispersion—where the variance exceeds the mean—due to hardware fluctuations or complex signal environments. When overdispersion is present, Poisson-based models often yield blurry or artifact-heavy reconstructions.

This paper proposes a more robust framework for reconstructing biomedical images from low-photon counts by replacing the standard Poisson model with a Negative Binomial (NB) distribution. By combining the NB model with $\ell_p$ Total Variation ($0 < p < 1$) regularization, we aim to achieve sharper, more accurate reconstructions in high-noise, overdispersed scenarios.

Methodology

The reconstruction task is formulated as a non-convex optimization problem. The approach integrates three core components:

  1. Negative Binomial Likelihood: Instead of minimizing a Poisson-based loss, we minimize the negative log-likelihood of the NB distribution. The NB model is characterized by a dispersion parameter ($r$), where lower values of $r$ indicate higher overdispersion.
  2. $\ell_p$ Total Variation (TV) Penalty: To promote sparsity and preserve sharp edges, we employ the $\ell_p$ TV quasi-seminorm ($0 < p < 1$). This non-convex penalty is known to outperform the standard $\ell_1$ TV in edge preservation. We investigate both isotropic (rotation-invariant) and anisotropic variants.
  3. Optimization Strategy:
  4. Second-Order Approximation: The complex objective function is approximated using second-order Taylor expansions to create simpler quadratic sub-problems.
  5. Barzilai-Borwein (BB) Method: To manage computational costs, the BB method is used to approximate the Hessian matrix with a scalar value.
  6. Iteratively Reweighted $\ell_1$: The non-convex $\ell_p$ term is handled by transforming it into a sequence of weighted $\ell_1$ problems, solvable via standard gradient projection methods.

Experimental Setup

The study utilized 2D biomedical images (lung CT scans) with dimensions of $512 \times 512$ pixels from the MATLAB Medical Imaging Toolbox. To simulate overdispersed noise, images were subjected to Gaussian blur and then sampled from a Negative Binomial distribution with varying dispersion parameters ($r = 1, 10, 25, 1000$).

Original and noisy simulated images at various dispersion levels.

Results and Discussion

Quantitative analysis focused on the Root Mean Square Error (RMSE) across different noise levels and $p$-values.

Performance and Stability

A key finding was the stability of reconstruction accuracy across different values of $p$ (from 0.1 to 0.9). As shown in the performance plots, the RMSE remains remarkably consistent for a given noise level $r$, regardless of the specific $p$-value chosen. This suggests that the model is robust and does not require extensive fine-tuning of the $\ell_p$ power.

RMSE performance across p-values and dispersion levels for anisotropic and isotropic penalties.

NB vs. Poisson Models

In high-noise scenarios ($r=10$), the NB model achieved an RMSE of 13.96%, significantly outperforming the Poisson model's 15.35%. As the noise decreases ($r=1000$), the performance gap vanishes, as the NB distribution mathematically converges to the Poisson distribution. Qualitatively, the NB reconstruction retains significantly more structural detail and contrast in lung tissue compared to Poisson-based approaches.

Conclusion

The Negative Binomial model combined with $\ell_p$ TV regularization provides a superior framework for low-photon count imaging under overdispersion. Future work will focus on implementing automatic estimation of the dispersion parameter $r$ from raw data, accelerating the non-convex optimization for real-time applications, and extending the framework to 3D volumetric data.

Summary updated

Abstract

The negative binomial model, which generalizes the Poisson distribution model, can be found in applications involving low-photon signal recovery, including medical imaging. Recent studies have explored several regularization terms for the negative binomial model, such as the ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</inf> quasi-norm with 0 < p < 1, ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> norm, and the total variation (TV) quasi-seminorm for promoting sparsity in signal recovery. These penalty terms have been shown to improve image reconstruction outcomes. In this paper, we investigate the ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</inf> quasi-seminorm, both isotropic and anisotropic ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</inf> TV quasi-seminorms, within the framework of the negative binomial statistical model. This problem can be formulated as an optimization problem, which we solve using a gradient-based approach. We present comparisons between the negative binomial and Poisson statistical models using the ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</inf> TV quasi-seminorm as well as common penalty terms. Our experimental results highlight the efficacy of the proposed method.
BibTeX
@misc{lu2024sparse,
  author = {Yu Lu and Roummel F. Marcia},
  title = {{Sparse Signal Reconstruction for Overdispersed Low-Photon Count Biomedical Imaging Using ℓ<sub>p</sub> Total Variation}},
  year = {2024},
  eprint = {2408.16622},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.16622},
}
Figure