Acoustic-Impedance-Aware Ultrasound NeRF

• The paper introduces a neural rendering framework that encodes spatial acoustic impedance into NeRF for precise 3D ultrasound imaging.
• It employs a dual-branch network with multi-resolution hash-grid and spherical harmonics to model tissue physics and view-dependent ultrasound reflections.
• Results demonstrate significant improvements in reconstruction fidelity, inference speed, and probe localization accuracy over traditional and prior NeRF models.

Acoustic-Impedance-Aware Ultrasound NeRF (AIA-UltraNeRF) is a neural volume rendering framework for 3D ultrasound imaging that encodes spatial acoustic impedance directly into neural radiance fields (NeRF) through multi-resolution hash encodings. It leverages a physically informed forward model of ultrasound wave propagation, enabling high-fidelity reconstruction and rapid localization of probe positions for robotic ultrasound systems. The approach fully decouples data acquisition from interpretation workflows, enabling operator-independent and efficient 3D ultrasound diagnostics with significant speed and accuracy enhancements over traditional and prior NeRF-based pipelines (Zhang et al., 23 Nov 2025).

1. Acoustic Impedance and Forward Model

AIA-UltraNeRF encapsulates the essential physics of ultrasound by explicitly modeling local acoustic impedance, a tissue property defined as $Z(x) = \rho(x) \cdot c(x)$, where $\rho(x)$ denotes the local mass density and $c(x)$ the speed of sound. At an interface between samples $x_1$ and $x_2$, energy reflectivity is determined by

$$R_{\mathrm{phys}}(x_1 \rightarrow x_2) = \left| \frac{Z(x_2) - Z(x_1)}{Z(x_2) + Z(x_1)} \right|^2.$$

The observed intensity at the transducer aggregates exponentially attenuated reflections and backscatter along a ray $r(t) = o + t \cdot d$, via a volume-rendering integral: $$I_{\mathrm{rec}}(d) = \int_0^D R_{\mathrm{eff}}(r(t), d) \cdot I(t)\, dt,$$ where $I(t) = I_0 \exp(-\int_0^t \alpha(r(\tau)) d\tau)$ accounts for cumulative attenuation $\alpha(x)$ and $\rho(x)$0 combines reflectivity, angular dependence, and scattering terms. This formulation ensures that the reconstructed signal conforms to underlying acoustic physics at both macroscopic and interface levels (Guo et al., 2024).

2. Network Architecture and Hash-Grid Impedance Encoding

AIA-UltraNeRF employs a two-branch architecture integrating multi-layer perceptrons (MLPs) with hash-grid encodings:

1. Spatial Acoustic Encoder:

The continuous function $\rho(x)$1 models the 3D volume as $\rho(x)$2 where spatial coordinates $\rho(x)$3 are encoded into multi-resolution hash grids that implicitly store local acoustic impedance $\rho(x)$4. Feature vectors $\rho(x)$5 are computed by hashing integer triplets and trilinear interpolation across levels, producing a concatenated feature $\rho(x)$6.

2. Directional Branch: The directional dependence of ultrasound reflections is encoded via spherical harmonics $\rho(x)$7 for ray direction $\rho(x)$8. This enables view-dependent estimation of B-mode intensity $\rho(x)$9. The full system predicts volumetric density $c(x)$0 and a latent vector $c(x)$1 through MLP$c(x)$2, and subsequently the per-ray color via MLP$c(x)$3.

This hash-grid scheme allows for sparse, high-dimensional acoustic information storage, significantly reducing the sampling requirements of previous continuous NeRF models, thereby accelerating both training and inference (Zhang et al., 23 Nov 2025).

3. Volume Rendering with Impedance Integration

For any pixel $c(x)$4, a ray $c(x)$5 is cast, and the final intensity is synthesized by integrating view-dependent B-mode responses modulated by local density, acoustic impedance, and direction: $c(x)$6 with

$c(x)$7

Both density $c(x)$8 and output $c(x)$9 depend on hash-encoded impedance and directional encoding, capturing angular variation of interface reflectivity and scattering. This model directly reflects the physics of real tissue in ultrasound—attenuation, reflection governed by impedance mismatches, and position-dependent backscatter (Guo et al., 2024, Zhang et al., 23 Nov 2025).

4. Dual-Supervision Hashing for Fast Localization

AIA-UltraNeRF introduces a dual-supervised teacher–student framework for rapid probe localization:

• Teacher receives weakly augmented rendered images, producing hash codes $x_1$0.
• Student receives strongly augmented variants, producing $x_1$1.
• The hashing network is trained with a combination of three losses:
  • Self-distilled hashing loss:

  $x_1$2 - Hybrid proxy-label loss: $x_1$3 using learnable proxies and cross-entropy over class scores - Quantization loss: Encouraging hash components towards $x_1$4 using a Gaussian estimator

• The total localization loss is

$x_1$5

Offline, a large database of rendered images with known poses is precomputed, enabling near-instantaneous pose retrieval via hash code comparison, rather than repeated rendering. This decouples image reconstruction from real-time localization and allows efficient robotic probe refinement only when necessary (Zhang et al., 23 Nov 2025).

5. Robotic System Integration and Scanning Modes

AIA-UltraNeRF is deployed within a robotic ultrasound system (RUSS) that features a 7-DoF serial arm and a compact 3-axis spherical remote center-of-motion (RCM) mechanism. The RCM kinematics ensure the probe tip rotates purely about a fixed center beneath the skin ($x_1$6), with controllable joints for azimuth, tilt, and spin. This enables both annular scans (holding tilt, rotating azimuth through $x_1$7 for phantoms) and fixed-point rotations (spinning at fixed pitch/yaw for humans).

All 2D frames are acquired with precise spatio-angular metadata, allowing fully offline construction of the 3D ultrasound map via AIA-UltraNeRF. This separation of acquisition from interpretation streamlines clinical workflows and supports operator-independent imaging (Zhang et al., 23 Nov 2025).

6. Quantitative Performance and Comparative Results

Empirical evaluations demonstrate substantial performance improvements over baseline NeRF models and instant neural graphics primitives (NGP):

• Phantom Reconstruction (PSNR/SSIM/LPIPS):

  • Vanilla NeRF: 17.6842 / 0.6788 / 0.4271
  • Instant-NGP: 20.1199 / 0.7213 / 0.4050
  • AIA-UltraNeRF: 20.2560 / 0.7569 / 0.2980
• Human Subjects (10 people) (PSNR/SSIM/LPIPS):
  • Instant-NGP: 22.1358 / 0.7920 / 0.3154
  • AIA-UltraNeRF: 22.7414 / 0.8168 / 0.2881
• Inference Speed: $x_1$82 s/image (AIA-UltraNeRF) vs. $x_1$922 s/image (vanilla NeRF); $x_2$0 speedup.
• Localization Accuracy: Mean errors of 3.07°–4.03° (ViT-Base, DeiT-Base, ResNet50 backbones); most intra- and cross-subject queries localize within 0–10° error; retrieval SSIM is inversely correlated with pose error.

These improvements arise from the impedance-aware hash encoding, physics-guided rendering, and the decoupled localization framework (Zhang et al., 23 Nov 2025).

7. Extensions, Best Practices, and Context

Recommended best practices for AIA-UltraNeRF and related architectures include pretraining the spatial MLP on prior CT/MRI-derived impedance maps, employing sine-activation-based MLPs (SIREN) for high-frequency detail, and accommodating advanced phenomena such as ray-bending due to strong $x_2$1 variation or reverberation via secondary ray tracing. The volume-rendering formulation is extensible: hierarchical sampling (coarse-fine passes), adversarial losses for texture, and multi-bounce modeling are supported. AIA-UltraNeRF’s explicit impedance integration addresses the key deficiency in prior ultrasound NeRF works—namely, the neglect of tissue-dependent signal characteristics essential for diagnostic value (Guo et al., 2024, Zhang et al., 23 Nov 2025).

A plausible implication is that acoustic-impedance-aware neural rendering will become critical for future robotic and automated ultrasound diagnostics, facilitating hardware–software co-design and analytic labeling of tissue boundaries, structure, and pathology.