Joint Reconstruction of Absorbed Optical Energy Density and Sound Speed Distributions in Photoacoustic Computed Tomography: A Numerical Investigation

Chao Huang, Kun Wang, Robert W. Schoonover, Lihong V. Wang, Mark A. AnastasioIEEE Trans. Computational Imaging 2016 · DOI

TL;DR

This paper investigates the challenging problem of simultaneously reconstructing an object's optical absorption and sound speed distribution from photoacoustic computed tomography (PACT) data. The authors develop a numerical framework to explore the feasibility and limitations of this "joint reconstruction" (JR) approach, revealing that while promising, it faces significant challenges due to the ill-posed nature of the problem.

What Problem Does This Paper Address?

Photoacoustic computed tomography (PACT) is a powerful imaging technique that uses light absorption to generate ultrasound waves, which are then detected to reconstruct an image of the object's internal optical energy deposition. Traditional PACT methods assume a uniform speed of sound (SOS) in the imaging medium. However, variations in SOS can cause significant wavefield distortions, leading to artifacts and reduced image quality. While incorporating knowledge of the SOS distribution can improve PACT images, accurately estimating this distribution beforehand is difficult and often requires a separate ultrasound computed tomography (USCT) study.

This paper addresses the question of whether it's possible to simultaneously reconstruct both the optical absorption (A(r)) and the SOS distribution (c(r)) directly from the PACT measurements themselves. This "joint reconstruction" (JR) problem aims to circumvent the need for a separate USCT scan. However, the practical feasibility and numerical properties of this JR problem are not well understood, which is the gap this paper aims to fill through numerical investigation.

Methods

The authors propose an alternating optimization approach to solve the JR problem. This strategy breaks down the complex JR problem into two simpler sub-problems that are solved iteratively:

  1. Reconstruction of A(r) given c(r): This sub-problem is solved using a previously developed full-wave iterative method, incorporating a total variation (TV) penalty for regularization.
  2. Reconstruction of c(r) given A(r): This sub-problem is formulated as a nonlinear optimization problem. The authors derive the Fréchet derivative of the objective function with respect to c(r) using an adjoint method. This gradient is then used by a limited-memory BFGS (L-BFGS) algorithm to find the optimal c(r).

The overall JR algorithm alternates between these two sub-problems until convergence is achieved. The paper also uses a discrete imaging model based on the wave equation to simulate PACT measurements.

Datasets

The paper does not use real-world datasets. Instead, it relies on simulated photoacoustic (PA) measurement data generated from numerical phantoms representing the optical absorption (A(r)) and speed of sound (c(r)) distributions. These phantoms are designed to mimic realistic scenarios and explore specific properties of the JR problem, such as the effects of spatial supports, spatial bandwidths, and perturbations in A(r).

Key Insights and Ideas

  • Non-convexity of the JR Cost Function: The data fidelity term of the JR cost function is shown to be non-convex with respect to the SOS distribution (c(r)). This implies that the optimization problem can have multiple local minima, making it challenging to find the true global minimum and suggesting that accurate initial estimates for c(r) are crucial.
  • Support Conjecture: While a theoretical "support conjecture" (where the support of c(r) must be enclosed within the support of A(r)) was proposed for accurate reconstruction under geometrical acoustics, numerical results suggest that this conjecture is not strictly necessary for accurate reconstruction when using a full-wave model, especially with perfect data. However, the overlap between the supports of A(r) and c(r) does influence the ability to solve the JR problem.
  • K-Space Conjecture: The relative spatial frequency bandwidths of A(r) and c(r) significantly impact the numerical stability of Sub-Problem #2 and the overall JR. For accurate reconstruction, especially in the presence of noise, the spatial frequency bandwidth of A(r) should be comparable to or larger than that of c(r).
  • Ill-Conditioned Nature: The JR problem is inherently ill-conditioned. Small perturbations in the assumed A(r) can lead to large changes in the reconstructed c(r), and vice-versa. This sensitivity suggests that accurate JR might not always be possible, especially with imperfect data or modeling.
  • Regularization is Key: The numerical instability of the JR problem can be mitigated by incorporating appropriate regularization techniques (e.g., total variation for A(r) and other penalty functions for c(r)). However, even with perfect data, regularization is often required to obtain reasonable estimates.

Key Results

  • The paper demonstrates that the joint reconstruction of A(r) and c(r) is numerically challenging due to the non-convexity and ill-conditioned nature of the problem.
  • Numerical experiments show that while the "support conjecture" is not a strict requirement for full-wave models, the overlap of supports between A(r) and c(r) plays a role.
  • The "k-space conjecture" is proposed: the spatial frequency bandwidth of A(r) should be comparable to or larger than that of c(r) for stable reconstruction, particularly in noisy conditions.
  • The study shows that even with perfect, noiseless simulated data and phantoms satisfying theoretical conjectures, accurate JR is difficult without regularization.
  • Regularization significantly improves the accuracy of reconstructed images, making JR more robust to noise and modeling errors.
  • The JR approach, when successful, can yield more accurate estimates of A(r) compared to standard PACT methods that assume a constant SOS.

Limitations and Future Work

  • Numerical Instability: The primary limitation highlighted is the inherent numerical instability of the JR problem, which makes accurate reconstruction challenging.
  • Idealized Model: The study primarily uses a lossless fluid medium and neglects certain physical effects like acoustic attenuation and transducer response imperfections in some simulations. While some simulations explore these, a comprehensive treatment is complex.
  • Computational Cost: The 2D formulation is used for computational convenience, but extending the full-wave model to 3D JR is computationally intensive.
  • Future Work: The authors suggest that investigating the JR problem with experimental data is a crucial next step. They also propose relaxing the assumptions of a lossless medium and exploring the incorporation of additional information, such as ultrasound computed tomography (USCT) measurements, to improve the feasibility of JR. The computational burden of 3D JR and the need for accurate modeling of the imaging operator are also identified as areas for future research.

Summary updated

BibTeX
@article{huang2016joint,
  author = {Chao Huang and Kun Wang and Robert W. Schoonover and Lihong V. Wang and Mark A. Anastasio},
  title = {{Joint Reconstruction of Absorbed Optical Energy Density and Sound Speed Distributions in Photoacoustic Computed Tomography: A Numerical Investigation}},
  journal = {IEEE Trans. Computational Imaging},
  year = {2016},
  doi = {10.1109/tci.2016.2523427},
}
Figure