Persistent-Homology-Guided Topology Scanning of Qualitative Indicators for Acoustic Inverse Scattering

This paper proposes a topology-aware postprocessing framework that utilizes persistent homology to automatically determine optimal thresholds for qualitative acoustic inverse scattering indicators, thereby enabling robust reconstruction of scatterers with complex topologies like multiple components or holes.

The Gist

Imagine you are trying to find hidden objects in a dark room using a special flashlight. This flashlight doesn't show you a clear picture of the objects; instead, it paints a fuzzy, gray-scale map on the wall. The brighter the spot on the map, the more likely it is that an object is there.

This is the problem scientists face in acoustic inverse scattering (like using sound waves to "see" inside the body or the ground). They get these fuzzy maps, called indicators, but they need a clear, black-and-white picture to know exactly where the objects are and what shape they have.

The Problem: The "Threshold" Trap

To turn the fuzzy gray map into a clear black-and-white picture, you have to draw a line. You say, "Anything brighter than this gray level is an object; anything darker is empty space."

In the past, scientists had to guess this line (the threshold) by eye. This was risky:

• If the line was too low, the map might show ghost objects (tiny specks of noise looking like separate islands).
• If the line was too high, it might eat holes (missing the empty space inside a ring-shaped object) or break a single object into several pieces.

This is especially tricky if the hidden object has a complex shape, like a donut (which has a hole) or two separate islands.

The Solution: The "Topological Detective"

This paper introduces a new method called Persistent Homology. Think of this as a topological detective that doesn't just look at one gray level, but watches the map change as you slowly turn up the brightness.

Here is how the detective works, using a simple analogy:

1. The Water Level Analogy: Imagine the gray map is a landscape where high values are tall mountains and low values are valleys.

   • The Detective's Job: Instead of picking one water level to flood the map, the detective slowly raises the water from the bottom up.
   • Tracking Islands (H0): As the water rises, new islands (connected components) appear. Some islands are tiny and get swallowed by the water immediately (these are likely noise). Others are big mountains that stay above water for a long time. The detective ignores the tiny, fleeting islands and counts only the long-lasting ones.
   • Tracking Lakes (H1): As the water rises, it might fill up a valley to create a lake (a hole in the island). Some lakes are just puddles that fill up instantly. Others are deep lakes that stay open for a long time. The detective counts only the deep, persistent lakes.

2. The "Lifetime" Clue: The detective measures the lifetime of each island and lake.

   • Short life: "This island appeared and disappeared quickly. It's probably just a glitch or noise." -> Ignore it.
   • Long life: "This island has been here from the start and is still here. This is a real object." -> Keep it.

How the New Method Works

Once the detective has counted the "real" islands and lakes, the paper proposes a two-step process:

1. Count the Features: The method looks at the "persistence diagrams" (a chart of lifetimes) to decide: "Okay, the real object probably has 2 separate parts and 1 hole."
2. Find the Perfect Line: Now, instead of guessing the gray level, the computer scans through all possible levels. It stops at the exact level where the resulting black-and-white picture matches the detective's count (2 parts, 1 hole) and isn't too big or too small.

Why This Matters

The paper tested this on three different types of "flashlights" (mathematical indicators):

• The "Noisy" Flashlight: When the map was very messy, the old method (guessing the line) broke the objects into pieces and missed the holes. The new method fixed this, correctly identifying the shape and the hole.
• The "Clean" Flashlight: When the map was already clear, the new method didn't mess things up; it just confirmed the shape was correct.

The Bottom Line

This paper doesn't invent a new flashlight; it invents a smarter way to develop the photo. By using Persistent Homology (the detective that tracks lifetimes), the method automatically figures out the correct number of objects and holes, ensuring the final image is topologically correct (e.g., it knows a donut has a hole, and it doesn't accidentally split one object into two).

It works with any existing sound-scattering method and turns fuzzy, noisy data into a reliable, shape-aware reconstruction.

Technical Summary

Technical Summary: Persistent-Homology-Guided Topology Scanning for Acoustic Inverse Scattering

Problem Statement
Qualitative methods for acoustic inverse scattering, such as the Linear Sampling Method (LSM) and the Factorization Method (FM), reconstruct scatterers by generating indicator functions defined on a sampling window. These indicators are typically gray-scale fields where high values suggest the presence of a scatterer. To obtain a binary reconstruction of the obstacle, a threshold must be applied. However, this thresholding step is often empirical and unstable. In the presence of noise-induced artifacts or when the scatterer possesses nontrivial topology (e.g., multiple components or interior holes), simple thresholding can lead to incorrect topological reconstructions, such as fragmenting connected components or failing to recover holes.

Methodology
The paper proposes a topology-aware postprocessing framework based on persistent homology to guide the thresholding process. The methodology proceeds as follows:

1. Indicator Conversion: A normalized qualitative indicator $I$ (ranging from 0 to 1) is converted into a distance-type indicator $\rho = 1 - I$. Lower values of $\rho$ correspond to higher evidence of the scatterer.
2. Persistent Homology Scan: The method applies cubical persistent homology to the sublevel sets of $\rho$. As the level $s$ increases, the sets $D_s = \{z : \rho(z) \leq s\}$ form a filtration. Persistent homology tracks the birth and death of topological features (connected components in $H_0$ and holes in $H_1$) across this filtration.
3. Topology Estimation or Prescription:
   • Unknown-Topology Mode: The method estimates the target topology by analyzing the persistence diagrams. It identifies "dominant" features based on their lifetimes (death time minus birth time) and the gaps between them. The number of connected components is estimated as $1 + N_0$ (where $N_0$ is the count of dominant finite $H_0$ bars), and the number of holes is estimated as $N_1$ (the count of dominant $H_1$ bars).
   • Known-Topology Mode: If prior information is available, the target Betti numbers $(\beta^*_0, \beta^*_1)$ are prescribed directly.
4. Topology-Guided Threshold Selection: The algorithm scans candidate levels $s$ to find an optimal reconstruction level $s^*$. The selection minimizes a score function $J(s)$ composed of:
   • Topological Discrepancy: The absolute difference between the Betti numbers of the candidate reconstruction and the target topology.
   • Geometric Penalties: Mild penalties for reconstructions that are too small, too large, or deviate from a preferred area fraction.
   • Confidence Term: A term favoring regions with higher mean indicator values.
5. Reconstruction: The final binary scatterer is defined as the sublevel set at the optimal level $s^*$. Spatial localization of components and holes is achieved via connected-component labeling on the resulting mask and its complement.

Key Contributions

• Topology-Aware Postprocessing: The paper introduces a framework that explicitly incorporates topological constraints into the thresholding of qualitative indicators, addressing the instability of empirical thresholds in complex geometries.
• Indicator Agnosticism: The method is designed to be independent of the specific qualitative indicator used. It can be applied to LSM, FM, or fused indicators, provided they are normalized.
• Automatic Topology Detection: It provides a rule-based mechanism to estimate the number of scatterer components and holes directly from the persistence lifetimes of the indicator field, utilizing lifetime gaps to distinguish robust features from noise.
• Mathematical Stability: The authors establish that the persistence diagrams of the distance-type indicator are stable with respect to $L^\infty$ perturbations of the input indicator. They further prove that the counting of dominant topological features remains stable if a sufficient gap exists in the persistence spectrum.

Results
Numerical experiments were conducted using a single-frequency setup ($k=2$) with 5% additive noise on synthetic data. The true scatterer consisted of one solid disk and one annular component (Target Topology: $\beta_0=2, \beta_1=1$).

• LSM Indicator: Standard fixed-level thresholding ($s=0.5$) failed to recover the correct topology, yielding $\beta=(4,0)$ due to fragmentation. The PH-guided method successfully corrected this, recovering the target topology $(2,1)$ and improving the Intersection-over-Union (IoU) score from 0.432 to 0.570.
• FM Indicator: The fixed-level threshold already produced the correct topology. The PH-guided method maintained this correctness while slightly improving the IoU (0.727 to 0.740), acting as a mild refinement rather than a drastic correction.
• Fused Indicator: Similar to FM, the fused indicator yielded correct topology with fixed thresholding, and the PH-guided method provided negligible geometric improvement, confirming that the core benefit lies in the topology-aware selection rather than the fusion itself.
• Persistence Analysis: The experiments demonstrated that the FM indicator exhibited clearer separation in persistence lifetimes compared to LSM, making automatic topology detection more robust for FM.

Significance and Claims
The paper claims that the primary significance of this work is the ability to transform qualitative gray-scale indicators into topology-consistent binary reconstructions. The method is particularly effective when standard thresholding fails to capture the correct topological structure of the scatterer. By leveraging the stability of persistent homology, the framework offers a systematic way to handle noise and artifacts without requiring multi-frequency data or complex regularization schemes. The authors emphasize that the method remains close to the classical single-frequency qualitative inverse scattering setting while adding a robust, automated layer for topological inference.