Efficient Inversion of Multiple-Scattering Model for Optical Diffraction Tomography

Emmanuel Soubies, Thanh-An Pham, Michael UnserarXiv 2017 · DOI · OpenReview

Abstract: Optical diffraction tomography relies on solving an inverse scattering problem governed by the wave equation. Classical reconstruction algorithms are based on linear approximations of the forward model (Born or Rytov), which limits their applicability to thin samples with low refractive-index contrasts. More recent works have shown the benefit of adopting nonlinear models. They account for multiple scattering and reflections, improving the quality of reconstruction. To reduce the complexity and memory requirements of these methods, we derive an explicit formula for the Jacobian matrix of the nonlinear Lippmann-Schwinger model which lends itself to an efficient evaluation of the gradient of the data-fidelity term. This allows us to deploy efficient methods to solve the corresponding inverse problem subject to sparsity constraints.

Introduction: The Challenge of Multiple Scattering

Optical Diffraction Tomography (ODT) is a powerful tool for imaging biological samples without the need for toxic dyes or staining. It works by measuring how light scatters as it passes through a sample to map its refractive index.

Schematic of the optical diffraction tomography setup showing the incident wave, the sample, and the resulting scattered waves.

Historically, researchers used linear approximations (like the Born or Rytov approximations) to solve this. However, these "shortcuts" fail when samples are thick or have high refractive index contrasts because they ignore multiple scattering—where light bounces off multiple structures within the sample.

Recent nonlinear methods like SEAGLE can handle multiple scattering, but they are "memory hogs." They rely on an error-backpropagation strategy that requires storing every intermediate step of the forward simulation to calculate the gradient. For large 3D volumes, this quickly exceeds the RAM of even high-end workstations.

Methodology: Decoupling the Jacobian

The authors propose a more elegant mathematical approach to avoid the memory bottleneck:

  1. Explicit Jacobian Derivation: Instead of using backpropagation (which is essentially automatic differentiation through the solver's iterations), they derive a closed-form expression for the Jacobian matrix of the Lippmann-Schwinger operator.
  2. Decoupled Computation: Because they have an explicit formula, they can solve the forward scattering problem and the gradient (the "backward" pass) independently. This means they no longer need to store intermediate solver iterates.
  3. Solver Flexibility: While previous methods were tied to specific iterative solvers (like Nesterov’s Accelerated Gradient Descent), this new framework allows the use of any quadratic optimizer. The authors demonstrate that the Conjugate Gradient (CG) method is significantly faster than Nesterov's for this specific problem.
  4. Regularization: They solve the inverse problem using a Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) combined with Total Variation (TV) regularization and non-negativity constraints to ensure the reconstructed images are sharp and physically plausible.

Key Insights and Results

The "Aha" Moment

The Jacobian of the nonlinear scattering model actually has a structure very similar to the forward model itself. Specifically, computing the gradient requires inverting an operator $(I - \text{diag}(\mathbf{f})\mathbf{G}^H)$ that has the same size and form as the forward operator. By doing a bit more calculus upfront (deriving the Jacobian), you can trade a massive amount of memory for a small amount of extra computation. This trade-off is essential for moving ODT from 2D research toys to 3D clinical/biological applications.

Performance Benchmarks

The authors evaluated their method using simulated Shepp-Logan phantoms and real experimental data from the Institut Fresnel's public database.

  • Speed: On real experimental data, the proposed method was 15 times faster than SEAGLE (6 minutes vs. 93 minutes).
  • Memory: For a $256 \times 256$ reconstruction, the new method used ~337 MB of RAM compared to SEAGLE's ~460 MB. The paper predicts that for large 3D volumes (e.g., $512 \times 512 \times 256$), the memory difference would be the difference between a standard desktop and an impossible-to-source supercomputer.
  • Quality: The Signal-to-Noise Ratio (SNR) and visual quality of the reconstructions were virtually identical to the more expensive SEAGLE method, proving that the efficiency gains come with no "accuracy tax."

Limitations and Future Work

While the math is designed for 3D and the memory arguments are most compelling there, the experimental results in this specific paper are demonstrated on 2D slices. Like all ODT methods, it still suffers from the "missing cone" of information caused by the limited range of angles at which one can realistically illuminate a sample.

Future directions include extending this explicit Jacobian framework to other types of waves (like acoustics or microwaves) and integrating it into stochastic gradient descent schemes to further speed up massive 3D reconstructions.

Summary updated

Abstract

Optical diffraction tomography relies on solving an inverse scattering problem governed by the wave equation. Classical reconstruction algorithms are based on linear approximations of the forward model (Born or Rytov), which limits their applicability to thin samples with low refractive-index contrasts. More recent works have shown the benefit of adopting nonlinear models. They account for multiple scattering and reflections, improving the quality of reconstruction. To reduce the complexity and memory requirements of these methods, we derive an explicit formula for the Jacobian matrix of the nonlinear Lippmann-Schwinger model which lends itself to an efficient evaluation of the gradient of the data- fidelity term. This allows us to deploy efficient methods to solve the corresponding inverse problem subject to sparsity constraints.
BibTeX
@misc{soubies2017efficient,
  author = {Emmanuel Soubies and Thanh-An Pham and Michael Unser},
  title = {{Efficient Inversion of Multiple-Scattering Model for Optical Diffraction Tomography}},
  year = {2017},
  eprint = {1708.00745},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1708.00745},
}
Figure