Abstract: Neural operator learning as a means of mapping between complex function spaces has garnered significant attention in the field of computational science and engineering (CS&E). In this paper, we apply Neural operator learning to the time-of-flight ultrasound computed tomography (USCT) problem. We learn the mapping between time-of-flight (TOF) data and the heterogeneous sound speed field using a full-wave solver to generate the training data. This novel application of operator learning circumnavigates the need to solve the computationally intensive iterative inverse problem. The operator learns the non-linear mapping offline and predicts the heterogeneous sound field with a single forward pass through the model. This is the first time operator learning has been used for ultrasound tomography and is the first step in potential real-time predictions of soft tissue distribution for tumor identification in breast imaging.
Introduction
Ultrasound Computed Tomography (USCT) offers a non-ionizing alternative to X-ray mammography for breast cancer screening. However, generating accurate images requires solving a complex "inverse problem": reconstructing the underlying distribution of sound speeds (which indicates tissue density and potential tumors) from the time it takes for sound waves to travel through the breast (Time-of-Flight or TOF).
Traditionally, this reconstruction relies on iterative numerical methods. While these full-wave solvers account for complex physics, they are computationally demanding and slow, preventing real-time clinical feedback. Researchers from the University of Utah have proposed a solution using Tensorized Fourier Neural Operators (T-FNOs) to bypass these iterations, enabling high-accuracy reconstructions in under 1.5 seconds.
Methodology
The T-FNO Architecture
Unlike standard convolutional neural networks (like U-Nets) that operate on fixed pixel grids, Neural Operators are designed to learn mappings between infinite-dimensional function spaces. The authors utilized a T-FNO with 7.3 million parameters, 64 modes, and 32 hidden channels. This model learns to map a 2D "emitter-by-receiver" TOF field directly to a 2D spatial sound speed (SS) field.
Data Generation and Simulation
To train the model, the researchers generated a large-scale synthetic dataset using Gaussian Random Fields (GRFs) to simulate realistic variations in soft tissue. These fields were binned into discrete sound speed values representing tissue (1450–1550 m/s), water (1500 m/s), and skin (1580 m/s). The k-Wave MATLAB toolbox, a gold-standard pseudo-spectral solver, was used to simulate sound wave propagation and generate the corresponding TOF data on a 128x128 grid.
Results and Performance
Generalization vs. Memorization
A critical finding was the performance gap between the T-FNO and a standard U-Net baseline. While the U-Net achieved lower training error, it tended to memorize the data rather than learn the underlying physics. In contrast, the T-FNO achieved significantly lower testing error, demonstrating superior generalization to unseen data.
Quantitative Accuracy and Robustness
In "High Correlation" test cases with clean data, the T-FNO achieved a Testing Mean Squared Error (MSE) of $2.01 \times 10^{-2}$, outperforming the U-Net's $2.52 \times 10^{-2}$. The T-FNO also exhibited remarkable robustness; when 10% white noise was added to the signals, its performance remained stable, whereas the U-Net's error increased more significantly.
Computational Efficiency
The T-FNO approach turns a heavy physics-based optimization problem into a fast image-to-image translation task. A single reconstruction takes approximately 1.44 seconds on an Nvidia Titan RTX, a massive speedup compared to traditional iterative full-wave inversion which can take minutes or hours.
Discussion and Future Work
While these results are promising, the study utilized synthetic data and simplified physics (focusing solely on TOF). Real-world breast tissue contains complex "clutter" and continuous gradients of sound speed that were not captured in the GRF models.
Future work will aim to incorporate more comprehensive wave physics—such as attenuation and scattering—by exploring the inverse Helmholtz problem. Additionally, the researchers plan to validate the T-FNO's clinical utility using non-synthetic, real breast image data. This work represents a significant first step toward real-time, high-fidelity ultrasound imaging for tumor identification.
Summary updated