To image the inside of a biological cell, map the subsurface of the earth, or peer through dense tissue, researchers send waves—such as light, sound, or microwaves—into a medium and measure how they scatter. Historically, the mathematics behind reconstructing a 3D image from these scattered waves has relied on a crucial simplification: the single-scattering assumption.
Under linear models like the First Born or Rytov approximations, it is assumed that an incoming wave interacts with an object exactly once before traveling straight to the sensor. For example, in Quantitative differential phase contrast imaging in an LED array microscope, Lei Tian derived a 2D weak object transfer function by assuming small absorption and phase variations. This linearizes the relationship between the object's physical properties and the measured wave field. If every photon or acoustic wave only bounces once, back-calculating the object's structure becomes a straightforward, computationally efficient linear inverse problem.
However, a major paradigm shift is underway. Researchers are moving away from these restrictive linear approximations to embrace exact, nonlinear forward models that can account for multiple scattering—where waves bounce inside an object multiple times before escaping.
The Breakdown of Linear Models
While the single-scattering assumption works well for thin, nearly transparent biological samples, it fails completely in more complex environments. When an object is physically thick, densely packed, or possesses a high refractive index contrast, a wave does not just scatter once. It bounces from one part of the object to another, creating a cascade of secondary interactions.
As the authors of Compressive Imaging with Iterative Forward Models point out, traditional linear models fail to accurately reconstruct objects with large physical dimensions or high contrasts because they completely ignore these secondary interactions. When multiple scattering dominates, treating the system as linear introduces severe image artifacts or causes the reconstruction to fail entirely.
This limitation spans multiple imaging modalities:
- Ultrasound Tomography: In Full wave 3D inverse scattering transmission ultrasound tomography in the presence of high contrast, Bilal Malik et al. demonstrate that linear approximations break down when waves encounter high-contrast structures like bone, leading to severe geometric distortions.
- Diffractive Optics: Michael Quellmalz et al. note in Total Variation-Based Reconstruction and Phase Retrieval for Diffraction Tomography that the first-order Born approximation strictly limits reconstructions to weak scatterers with low refractive index contrasts.
- Random Media: In Imaging in Random Media by Two-Point Coherent Interferometry, Liliana Borcea explains that in high-density media, the relationship between the measured signal and the underlying geometry becomes fundamentally non-linear, rendering Born-based models ineffective.
Modeling the Full Physics via the Lippmann-Schwinger Equation
To overcome these limitations, modern computational imaging frameworks simulate the exact, nonlinear physics of wave propagation. The mathematical cornerstone of this approach is the Lippmann-Schwinger equation.
Instead of assuming the internal field within the object is simply equal to the incident wave, the Lippmann-Schwinger equation treats the total field as a coupled system. The field at any point inside the object depends on the incident wave plus the scattered waves originating from every other point inside the object.
By explicitly accounting for this internal coupling, researchers can accurately simulate how light propagates through highly scattering environments. For instance, in Three-Dimensional Optical Diffraction Tomography With Lippmann-Schwinger Model, Thanh-an Pham et al. developed a 3D forward model based on the exact, nonlinear Lippmann-Schwinger equation. By introducing numerical techniques to handle Green's function singularities, the authors enabled high-fidelity 3D refractive index reconstructions that significantly outperform traditional linear methods for thick or highly scattering samples.
Computational Strategies for Nonlinear Inversion
If the forward model is nonlinear, inverting it to reconstruct the image becomes a massive optimization challenge. How do we efficiently compute the gradient of a wave field that has bounced multiple times?
One powerful strategy is algorithm unrolling. In SEAGLE: Sparsity-Driven Image Reconstruction under Multiple Scattering, Hsiou-Yuan Liu et al. introduced a framework that unrolls Nesterov’s accelerated gradient descent directly into the forward model. By treating the iterations of the physics solver as layers in a computational graph, they enabled efficient backpropagation-style gradient calculations to reconstruct objects under strong multiple scattering.
To accelerate this process further, Yanting Ma et al. introduced CISOR in Accelerated Image Reconstruction for Nonlinear Diffractive Imaging. CISOR utilizes a relaxed version of the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) to provide memory-efficient, stable reconstructions of high-permittivity objects without taking prohibitive amounts of time.
Linear (Born) Approach:
Incident Wave ───> [ Object Layer ] ───> Single Scatter ───> Sensor (Inaccurate for thick media)
Nonlinear (Unrolled) Approach:
Incident Wave ───> [ Internal Coupling / Multiple Bounces ] ───> Sensor
▲ │
└──── Gradient Feedback ──┘ (Iterative Physics Solver)
Other approaches tackle the problem by adapting the propagation physics to non-paraxial regimes. Jiabei Zhu et al. introduced a split-step non-paraxial multiple-scattering model in High-fidelity intensity diffraction tomography with a non-paraxial multiple-scattering model., showing that accounting for both multiple scattering and wide-angle light propagation is essential for high-fidelity 3D imaging of dense biological tissues under high numerical aperture illumination.
Expanding the Horizons of Tomography
Moving beyond the weak scattering assumption opens up new capabilities across diverse imaging modalities:
- Chemical Imaging: In Bond-selective intensity diffraction tomography, Jian Zhao et al. demonstrated high-speed 3D chemical imaging of biological samples, noting that integrating advanced multiple-scattering models will be key to expanding the technique from single cells to thick tissues.
- Multi-Modal Imaging: In Three-dimensional bi-functional refractive index and fluorescence microscopy (BRIEF), Yi Xue et al. developed a technique to simultaneously reconstruct 3D refractive index and fluorescence, using the physical model to digitally correct scattering artifacts in fluorescence images.
- Dynamic Targets: Even when handling moving targets where phase information is completely missing, Michael Quellmalz et al. showed in Total Variation-Based Reconstruction and Phase Retrieval for Diffraction Tomography with an Arbitrarily Moving Object that combining regularization techniques like Total Variation with robust scattering models allows for high-quality 3D reconstructions.
Ultimately, by replacing the mathematical convenience of the single-scattering assumption with computationally unrolled, exact physical models, this paradigm shift enables high-resolution, artifact-free imaging deep inside complex, real-world media.
Key Foundational Work
Passive Array Correlation-Based Imaging in a Random Waveguide (2013) — contributed 1 concept elements to this direction.
Quantitative differential phase contrast imaging in an LED array microscope (2015) — contributed 1 concept elements to this direction.