UltraScatter: Ray-Based Simulation of Ultrasound Scattering

Basic Information

• Paper ID: 2510.10612
• Title: UltraScatter: Ray-Based Simulation of Ultrasound Scattering
• Authors: Felix Duelmer, Mohammad Farid Azampour, Nassir Navab
• Classification: physics.med-ph cs.CV
• Publication Date: October 12, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.10612

Abstract

Traditional ultrasound simulation methods achieve high accuracy by numerically solving the wave equation, but at enormous computational cost. Fast alternative methods based on convolution of precomputed impulse responses remain relatively slow, typically requiring several minutes to generate a complete B-mode image. This paper presents UltraScatter, a probabilistic ray tracing framework capable of efficiently and realistically modeling ultrasound scattering. Tissue is represented as volumetric fields of scattering probability and scattering amplitude, with ray interactions simulated through free-flight delta tracking. Scattered rays are traced to the sensor, with phase information integrated through a linear time-of-flight model. Combined with plane wave imaging and beamforming, the parallelized ray tracing architecture generates B-mode images within seconds. Validation using phantom data demonstrates realistic speckle and inclusion patterns, positioning UltraScatter as a scalable alternative to wave equation methods.

Research Background and Motivation

Problem Definition

Ultrasound simulation is crucial for enhancing reconstruction algorithms, optimizing sensor design, and training machine learning models under controlled conditions. However, existing methods face a trade-off between computational efficiency and physical realism.

Limitations of Existing Methods

1. Wave equation solvers (e.g., FDTD, k-space pseudospectral methods): High accuracy but prohibitive computational cost, difficult to scale
2. Fast methods such as Field II: Precompute spatial impulse responses through convolution, still requiring minutes to generate a single B-mode image
3. SIMUS: Based on geometric round-trip distances with delay-and-sum operations, limited efficiency

Research Motivation

There is a need for a method combining physical realism with fast computation to enable real-time or interactive ultrasound simulation. Physics-based rendering (PBR) techniques from computer graphics provide inspiration, particularly highly optimized CUDA-accelerated ray tracing algorithms.

Core Contributions

1. Modular high-performance framework: Models attenuation, absorption, and multiple scattering in participating media
2. Transmitter sampling strategy: Connects each scene interaction with all sensor elements
3. Complete transmit-receive beamforming chain: Directly converts simulated echoes to B-mode images
4. Significant performance improvement: Approximately 70-fold speedup compared to traditional frequency-domain implementations

Methodology

Task Definition

Starting from a labeled map, assign scattering characteristics to each tissue category. Use a Monte Carlo ray tracing scheme to model pressure wave emission, scattering, and attenuation. Write echoes returning to the sensor into RF buffers for specific elements, then generate the final B-mode image through conventional digital beamforming.

Model Architecture

Core Mathematical Model

The pressure signal P arriving at sensor element e at time t is defined as:

$P(e, t) = \int_\Omega \int_A P_i(x, t, \omega_i) f_d(\omega_i) d\omega da$

where $P_i(x, t, \omega_i)$ is the incident pressure from scene location x and direction $\omega_i$, and $f_d$ is the directivity function.

Scattered pressure is modeled as:

$P_{scattered}(x, t, \omega_o) = \int_\Omega a(x) p(\omega_i, \omega_o)P_i(x, t, \omega_i) d\omega_i$

where $a(x)$ represents scattering amplitude and $p(\omega_i, \omega_o)$ is the phase function controlling angular redistribution.

Ray Tracing Algorithm

1. Free-flight distance sampling: $s = s_{min} - \frac{1}{\mu} \ln(1-\xi)$

where $\mu$ is the maximum extinction coefficient throughout the medium and $\xi \sim U(0,1)$ is a random number.

2. Interaction determination: $\xi' < \frac{\sigma_t(x)}{\mu} \Rightarrow \text{true interaction, otherwise null interaction}$

3. Transmitter sampling strategy: For each scattering event, sample all sensor elements by emitting a set of secondary rays, each directed toward a sensor element.

4. Far-field approximation handling: Subdivide each sensor element into $\nu$ identical sub-elements: $\nu = \left\lceil \frac{2b}{\lambda_{min}} \right\rceil$

where $b$ is the azimuthal width of a single sensor element and $\lambda_{min}$ is the minimum wavelength in the transmission pulse.

Technical Innovations

1. Woodcock tracking: Employs free-flight delta tracking with null interactions to improve computational efficiency
2. Full-element transmitter sampling: Produces clearer images and sharper structural boundaries compared to selective methods
3. GPU parallelization: Based on the Mitsuba 3 framework, utilizing NVIDIA OptiX for efficient parallel computation
4. Phase information integration: Maintains phase consistency through a linear time-of-flight model

Experimental Setup

Dataset

Validation uses the CIRS General Purpose Phantom (Model 054GS), with its internal structure reconstructed according to manufacturer specifications and missing parameters adjusted until simulated references match clinical B-mode images.

Comparison Methods

1. Clinical scanner: Siemens Acuson Juniper with 12L3 linear array (192 elements, 2.9-11.5 MHz)
2. SIMUS: Frequency-domain time-harmonic simulator based on far-field and paraxial approximations

Implementation Details

• Center frequency: 6.2 MHz
• Imaging depth: 10 cm
• Dynamic range: 60 dB
• Sampling rate: 25 MHz
• F-number: 1.0
• 100,000 rays per sub-element
• Hardware: Intel Core i7-12700 CPU + NVIDIA RTX 4070 Ti GPU

Experimental Results

Main Results

Computational Performance Comparison:

• SIMUS: 634 ± 3 seconds
• UltraScatter: 9.3 ± 0.8 seconds
• Speedup: Approximately 70-fold

Image Quality Assessment:

1. Overall anatomical structure: All three methods (clinical, UltraScatter, SIMUS) share similar overall anatomical structures
2. Background speckle statistics: UltraScatter produces speckle statistics similar to real scans
3. Distal shadowing: Correctly reflects shadowing effects observed in real scans

Case Analysis

UltraScatter Characteristics:

• Exhibits over-illumination in the region directly below the probe, resulting from residual over-estimation of near-field gain
• Produces sharper inclusion boundaries, attributed to the full-element transmitter sampling strategy
• Provides clearer resolution of deep cylinders compared to SIMUS

SIMUS Characteristics:

• More faithfully reproduces the axial intensity attenuation of clinical systems
• Slightly less defined inclusion edges

Related Work

Traditional Ultrasound Simulation Methods

1. Full-wave solvers: FDTD, k-space pseudospectral methods - accurate but computationally expensive
2. Convolution methods: Field II - based on precomputed impulse responses
3. Geometric methods: SIMUS - based on delay-and-sum operations

Convolution Ray Tracing (CRT)

A recently emerged fast alternative that approximates acoustic wave propagation by tracing large ray ensembles with reflection and refraction at macroscopic boundaries, followed by modeling fine-scale scattering through convolution with separable point spread functions.

Physics-Based Rendering Techniques

PBR techniques from computer graphics, including ray marching, photon mapping, and delta tracking, provide highly optimized implementations for handling light propagation in heterogeneous media.

Conclusions and Discussion

Main Conclusions

UltraScatter successfully implements ray tracing ultrasound simulation, replacing frequency-domain solvers with Monte Carlo path sampling. Rendering typical CIRS phantom views requires approximately 9 seconds on the same hardware, compared to over 10 minutes for SIMUS, achieving significant performance improvement.

Limitations

1. Speckle statistics: Driven by random sampling, independent runs produce different patterns, with texture decorrelation as the probe moves laterally
2. Ray configuration: Current prototype emits rays from a single axial line to a cubic grid; more realistic configurations require full elevational aperture across sensors
3. Temporal coherence: Requires introduction of correlated random sequences or shared seed maps to maintain speckle stability during small probe motion

Future Directions

1. Elevational focusing: Achievable through dynamic receive delays or projecting additional rays
2. Enhanced physical modeling: Assign independent density, sound speed, and attenuation to each voxel
3. Nonlinear phenomena: Model macroscopic refraction, phase aberration, and even harmonic generation
4. Dynamic scene support: Improve speckle consistency and elevational resolution

In-Depth Evaluation

Strengths

1. Significant performance improvement: 70-fold speedup enables real-time simulation
2. Physical realism: Produces image quality and speckle patterns similar to clinical scans
3. Modular architecture: Design based on Mitsuba 3 facilitates extension and improvement
4. GPU parallelization: Fully leverages parallel computing capabilities of modern hardware
5. Complete pipeline: End-to-end solution from ray tracing to B-mode image generation

Limitations

1. Near-field effects: Over-estimation of near-field gain
2. Speckle consistency: Lacks temporal coherence, unsuitable for dynamic scenes
3. Dimensional constraints: Currently supports only 2D imaging, lacks complete 3D elevational processing
4. Validation scope: Validated only on a single phantom, lacks broader validation data

Impact

Academic Contributions:

• First successful application of computer graphics ray tracing techniques to ultrasound simulation
• Provides a new technical pathway for the ultrasound simulation field

Practical Value:

• Enables real-time ultrasound simulation for medical education and training
• Provides fast data generation tools for machine learning models
• Supports rapid iteration in sensor design optimization

Reproducibility:

• Based on open-source Mitsuba 3 framework
• Provides detailed implementation details and parameter settings
• Code and data availability requires further confirmation

Applicable Scenarios

1. Medical education: Real-time ultrasound simulation training systems
2. Algorithm development: Fast training data generation for deep learning models
3. Device design: Rapid prototype verification for sensor parameter optimization
4. Research applications: Scientific research projects requiring large amounts of simulation data

References

The paper cites 19 important references covering:

• Traditional ultrasound simulation methods (FDTD, k-Wave, Field II, SIMUS)
• Related work on ray tracing ultrasound simulation
• Computer graphics rendering techniques (Mitsuba 3, OptiX)
• Monte Carlo methods and delta tracking algorithms

Summary: UltraScatter represents a significant technical breakthrough in ultrasound simulation, successfully introducing advanced ray tracing techniques from computer graphics into medical ultrasound simulation, achieving a good balance between computational efficiency and physical realism. Despite some limitations, its significant performance improvement and modular design provide a solid foundation for future research.