Seeing Stiffness Behind Ultrasound Shadows

Shear wave elastography measures how stiff tissue is, which is how clinicians stage liver fibrosis and judge suspicious lumps. A focused ultrasound pulse pushes the tissue, which sends a sideways ripple (a shear wave) outward. The scanner tracks how fast that ripple travels, and faster ripples mean stiffer tissue. Wherever a rib, calcification, or gas pocket blocks the ultrasound beam, the scanner receives no echoes from the tissue behind it. That region becomes an acoustic shadow, and the stiffness map has a hole in it. The usual responses are to discard the hole or to smooth over it. Computer vision has a close analogue in disocclusion: regions hidden from the camera that a new view must fill in. The lesson from that literature is to separate what can be moved in from observed data from what must be invented. Geometry decides where the holes are. A generative model fills only what geometry cannot.

Ultrasound shadows differ from camera occlusions in one important way. The shadow hides tissue from the tracking beam, but the shear wave still travels through the hidden tissue and comes out the other side. The visible wavefield beyond the shadow therefore records the hidden stiffness, as delays and bends in the ripple. We propose to treat the shadow as a partially observable region. We would recover stiffness inside it by fitting the wave equation to the surrounding measurements. A per-pixel identifiability map would then separate the stiffness the data actually determine from the stiffness a learned prior merely supplies.

Core Claim

A frequency-domain wave inversion can return calibrated stiffness estimates inside acoustic shadows if it has three ingredients. First, it uses a continuous shadow-confidence map to down-weight untrustworthy displacement samples, rather than treating the shadow as missing data or as valid data. Second, it treats stiffness inside the shadow as an unknown governed by the same wave equation as the visible tissue. Third, it reports how strongly the outside data constrain each pixel inside. Such an inversion should beat interpolating the stiffness map and beat image-space inpainting, especially when a stiff inclusion sits partly or wholly inside the shadow. The claim that matters is the identifiability map. A clinician should be told which kilopascal values behind a rib are measured indirectly and which are guessed.

Relation to Prior Work

The method comes mainly from Plane Wave Elastography. It fits displacement data in the temporal frequency domain with the scalar wave (Helmholtz) equation, modeling each homogeneous patch as a sum of plane waves in arbitrary directions, so it needs no directional filtering. It assumes displacement is sampled everywhere in each patch. We keep its physics but let the data term vanish smoothly inside shadows while the wave-equation term stays active there.

Time-Harmonic Optical Flow constrains it. It tracks vibration amplitude with smoothness or total-variation regularizers. The research file reports that simple isotropic smoothing oversmooths or creates artificial peaks at heterogeneous boundaries. That is a direct warning against filling shadows with generic smoothness.

Weakly Supervised Estimation of Shadow Confidence Maps supplies the confidence map. It estimates continuous shadow confidence in real time (0.035 s per frame, Dice 0.7167) and shows that feeding these maps to downstream networks improves fetal biometrics by up to 7%. It never uses them inside a physical inversion, which is our use.

The disocclusion papers supply the design principle but not the tools. Synthesizing Images of Humans in Unseen Poses separates warping visible pixels from generating newly exposed ones. Coordinate-Based Texture Inpainting borrows real texture from visible regions instead of regressing blurry colors. Our version separates "propagated by physics" from "supplied by prior," and the wave equation plays the role their geometric warp plays.

AUE-Net is the opposite approach. It translates B-mode images into elastogram colors with a conditional adversarial network (SSIM 0.499 on thyroid). It cannot help inside a shadow, because the B-mode image there is also blank.

UltraGS models acoustic attenuation for B-mode view synthesis but not elasticity. It shows physics-aware rendering helps ultrasound, but it does not address our problem.

Research Program

We work with the complex displacement amplitude $u(x)$ at each excitation frequency $\omega$. We seek a shear modulus map $\mu(x)$ (stiffness in kPa) by solving

$$ \min_{u,\mu}\; \sum_x c(x)\,\lvert u(x)-u_{\text{obs}}(x)\rvert^2 \;+\; \lambda \sum_x \Big\lvert \nabla^2 u(x) + \tfrac{\rho\,\omega^2}{\mu(x)}\,u(x) \Big\rvert^2 \;+\; R(\mu), $$

where the symbols are defined as follows. The term $c(x)\in[0,1]$ is the shadow-confidence weight from the weakly supervised shadow network. The term $u_{\text{obs}}$ is the tracked displacement. The density $\rho$ is assumed constant. The weight $\lambda$ sets how strictly the wave equation is enforced. The regularizer $R$ encodes prior beliefs about stiffness.

Outside shadows, this objective behaves like a relaxed Plane Wave Elastography fit. Inside shadows, the data term fades, and $\mu$ is pinned only by the requirement that the wave crossing the region match what is seen on the far side. We would solve it by alternating updates: a linear least-squares step for $u$ and a nonlinear step for $\mu$.

The identifiability map comes from the curvature of the data-plus-physics terms with respect to $\mu$. We would compute it as the diagonal of a Gauss–Newton approximation to the posterior covariance. Pixels where this variance stays near the prior variance are labeled "prior-supplied." Pixels where it shrinks are labeled "wave-constrained."

We would compare two choices of $R$. One is total variation. The other is a learned prior: a small denoising diffusion model, which learns to turn noise into samples by repeated denoising. It would be trained on simulated piecewise-smooth stiffness maps and used as a plug-in regularizer. The learned prior is deliberately secondary. Its only job is to fill the non-identifiable part plausibly, and the output always marks where it did so.

Baselines

The naive reference is standard Plane Wave Elastography run on the shadowed data with uniform weights, as it is used today. We would score both its raw output inside the shadow and a version that simply masks the shadow, whose coverage inside the shadow is zero.

The strongest existing method on visible tissue is Plane Wave Elastography. On its noiseless digital phantom it reports a background modulus of 4.89 ± 1.27 kPa, an inclusion modulus of 14.76 ± 3.36 kPa, and a contrast-to-noise ratio (inclusion-versus-background separation relative to noise) of 8.78 dB. We would run it on the same phantom geometry with and without shadows. The unshadowed run is an upper bound, and our method should approach it inside the shadow. There is no reported result on Lacuna for any elastography method evaluated inside acoustic shadows, so the shadowed comparison is new.

Two hole-filling baselines are applied to Plane Wave Elastography's masked stiffness map. One is classical biharmonic interpolation. The other is MAT, an off-the-shelf image inpainting model (MAT), fine-tuned on the same simulated stiffness maps used to train our learned prior, so both methods see the same data.

The key ablation isolates the wave-through-the-shadow ingredient. It keeps the confidence-weighted objective but cuts the wave-equation term inside the shadow. Stiffness there is then filled by $R$ alone, which reduces to interpolation with a learned prior. If this ablation matches the full method, the central claim fails.

All methods would share the same patch size, frequency set, and noise level. Every method would get the same hyperparameter search budget, tuned on held-out phantoms. Beating the baselines means two things. Our method must cut the median relative stiffness error inside the shadow by a clear margin, for example to roughly half the interpolation error, when an inclusion overlaps the shadow. It must also stay statistically indistinguishable from interpolation when the shadowed tissue is homogeneous.

Why not just…?

Why not just interpolate the stiffness map across the shadow? For small shadows over homogeneous liver, this probably works about as well as anything. We expect to tie there and would report the tie. Interpolation cannot invent an inclusion it never saw, and smoothing-based fills misbehave at heterogeneous boundaries, as Time-Harmonic Optical Flow observes. The experiments place inclusions at controlled overlaps with the shadow to measure exactly where interpolation stops being adequate.

Why not just add the shadow-confidence weights to Plane Wave Elastography and stop? This is the most serious threat. Most of the gain on visible tissue may come from weighting alone. The inside-shadow ablation above separates these effects: weighting without the wave constraint inside the shadow cannot locate hidden inclusions. If it can, our physics component adds nothing and we would report that.

Why not just inpaint with a diffusion or image model, or translate B-mode to an elastogram? These approaches produce plausible textures with no physical anchor. AUE-Net's SSIM of 0.499 on unoccluded thyroid images already shows that image-space elastography is loose. Inside a shadow, the B-mode input that AUE-Net relies on is missing. The MAT baseline tests this directly, as does a check on whether its inpainted inclusions agree with ground truth or merely look right.

Why not just steer the beam or move the probe around the rib? Where an acoustic window exists, this is the better option and needs no algorithm. Rib spacing and bone often leave no such window, though. Our method is meant for that leftover case. It complements compounding rather than replacing it.

Proposed Experiments

The first stage is simulation. We would use the 40 mm × 40 mm inclusion phantom design from Plane Wave Elastography, run through an elastic wave simulator. We would impose shadow masks of varying width and depth, with realistic soft edges taken from shadow-confidence maps. We would vary how much an inclusion overlaps the shadow from none to complete. The metrics are relative modulus error inside the shadow, contrast-to-noise ratio for inclusions inside the shadow, and calibration: whether 90% posterior intervals contain the true modulus about 90% of the time in wave-constrained pixels.

The decisive check concerns an inclusion lying entirely inside the shadow. Our method should detect it from the outside field, while interpolation and the ablation cannot. Detection should fail exactly where the identifiability map predicts, for small or low-contrast inclusions far from the visible region. If detection fails where the map predicts success, the map is not trustworthy, and the method should not be used clinically.

The second stage uses physical CIRS phantoms like those in the Plane Wave Elastography paper. We would create real shadows by placing a bone-mimicking strip on the phantom surface, and use the same scan without the strip as ground truth.

The third stage is a modest retrospective check on clinical liver scans. We would compare stiffness recovered behind rib shadows against a nearby unshadowed intercostal acquisition of the same patient. The question is whether staging near the 10.45 kPa cutoff reported in the file changes when shadowed regions are included.