In many imaging fields—from medical ultrasound and seismic surveying to electron microscopy—the goal is to "see" the inside of an object by observing how waves bounce off or pass through it. This is an inverse problem: we have the data (the scattered waves) and we want to find the cause (the internal structure).
For decades, the standard approach has been to assume that waves interact with an object only once before reaching a sensor. This "single-scattering" assumption allows researchers to use linear mathematical models, like the Born or Rytov approximations. However, as Kim Knudsen, Jennifer J. Mueller, and Samuli Siltanen demonstrate in Reconstructions of piecewise constant conductivities by the D-bar method for electrical impedance tomography, these linear models fail when the object is "high-contrast"—meaning its properties differ significantly from its surroundings.
The emerging research direction across these 68 papers is the move toward nonlinear inverse scattering, which accounts for the full physics of wave propagation, including multiple scattering and curved paths.
The Linear Trap and Multiple Scattering
The core intuition behind the failure of linear models is the "circular dependency" of the wave field. In a simple linear model, we assume the wave inside the object is the same as the wave we sent in. In reality, as a wave enters a medium, the medium alters the wave; this altered wave then interacts with the next part of the medium, and so on.
When an object has high contrast (like bone in soft tissue or a metal pipe underground), waves undergo multiple scattering. A single photon or acoustic pulse might bounce between internal structures several times before exiting. As noted by Yu Sun and Ulugbek S. Kamilov in Deep Learning for Nonlinear Diffractive Imaging, ignoring these interactions leads to severe artifacts and a massive underestimation of the object's actual physical properties.
Mapping Curved Paths to Straight Lines
In Electrical Impedance Tomography (EIT), the problem is even more pronounced because electricity does not travel in rays; it follows the path of least resistance. This makes the relationship between boundary measurements and internal conductivity highly nonlinear.
To solve this, Melody Alsaker and Siiri Rautio introduced Virtual Hybrid Parallel-beam Tomography (VHPT) in CT scans without X-rays: parallel-beam imaging from nonlinear current flows. Their method mathematically transforms the curved, nonlinear current flows into "virtual radiographs"—straight-line projections similar to those used in CT scans. By decoupling the nonlinearity of the current flow from the reconstruction process, they achieve high-resolution imaging without the ionizing radiation of X-rays.
Overcoming Local Minima with Model Order Reduction
In seismic imaging, the nonlinearity of Full Waveform Inversion (FWI) often leads to "cycle skipping." Because waves are periodic, an optimization algorithm might mistakenly align the $n$-th peak of a predicted wave with the $(n+1)$-th peak of the observed data, getting trapped in a mathematically plausible but physically wrong "local minimum."
Researchers like Alexander V. Mamonov and Liliana Borcea have tackled this using Reduced Order Models (ROM). In Velocity estimation via model order reduction, they suggest that instead of matching raw waveforms, we should match the underlying mathematical operators (the ROMs) that describe the wave's evolution. This change of perspective makes the optimization problem nearly convex, allowing the algorithm to find the true subsurface velocity without getting lost in the "cycles."
Physics-Informed Deep Learning
The most recent shift involves using neural networks not just as "black box" image enhancers, but as solvers for the underlying physics. Traditional iterative solvers for nonlinear equations like the Lippmann-Schwinger equation are computationally expensive.
New frameworks are bridging this gap:
- Unrolled Optimizers: Zhou Chen and colleagues in MMV-Net "unroll" classical optimization steps into layers of a neural network, allowing the system to learn the best way to invert the nonlinear sensitivity matrix.
- Scattering Decoders: The ScaDec architecture, proposed by Yu Sun, replaces slow optimization with a U-Net that is specifically tested for stability against high permittivity contrasts, as detailed in Stability of Scattering Decoder For Nonlinear Diffractive Imaging.
- Full-Physics Tomography: In electron microscopy, Joel Yeo and Benedikt J. Daurer developed Ghostbuster, a method that replaces the linear "projection approximation" with a multislice forward model. This accounts for the curvature of the Ewald sphere and multiple scattering, effectively removing "ghost" artifacts in 3D reconstructions of thick biological samples (Ghostbuster: Beyond Ewald Sphere Curvature Correction).
Why It Works: The Multi-Frequency Advantage
A recurring theme in solving these nonlinear problems is the use of progressive refinement. Because low-frequency waves are less sensitive to fine details, they are less affected by the "chaos" of multiple scattering.
As explored by Owen Melia and Rebecca Willett in Multi-Frequency Progressive Refinement for Learned Inverse Scattering, starting with low-frequency data allows a model to capture the "big picture" (the bulk properties) first. Once the coarse structure is established, the model can safely incorporate higher frequencies to add sharp details without being misled by the complex interference patterns that high-contrast objects create.