Abstract
High-quality imaging in photoacoustic computed tomography (PACT) usually requires a high-channel count system for dense spatial sampling around the object to avoid aliasing-related artifacts. To reduce system complexity, various image reconstruction approaches, such as model-based (MB) and deep learning-based methods, have been explored to mitigate the artifacts associated with sparse-view acquisition. However, these methods often formulate the reconstruction problem in a discrete framework, making it prone to measurement and discretization errors; furthermore, the ill-posedness of the problem scales with the discretization resolution. In this work, an implicit neural representation (INR) framework is proposed for image reconstruction in PACT with ring transducer arrays. Specifically, the initial heat distribution is represented as a continuous function of spatial coordinates using a multi-layer perceptron (MLP). The weights of the MLP are determined through a self-supervised training process by minimizing the errors between measured and model-predicted photoacoustic signals. Simulation and phantom experiments demonstrate that the INR method excels in preserving image fidelity and suppressing artifacts compared to universal back-projection (UBP) and MB methods under the same acquisition conditions.
Introduction
Photoacoustic Computed Tomography (PACT) is a powerful "light-in, sound-out" imaging technique that uses laser pulses to induce thermoelastic expansion in tissue, emitting ultrasound waves captured by sensors. To achieve high-fidelity images, a dense array of sensors is typically required. However, high-channel count systems are expensive and complex. When engineers use fewer sensors—known as sparse-view acquisition—the resulting images suffer from "aliasing artifacts," which appear as streaks that can obscure critical anatomical details like small blood vessels.
Traditional reconstruction methods, such as Universal Back-Projection (UBP) and Model-Based (MB) iterative algorithms, operate on discrete pixel grids. This discretization introduces its own errors and makes the reconstruction problem increasingly difficult at higher resolutions. While supervised deep learning can help, it requires massive pre-existing datasets that are often unavailable in clinical settings. This paper proposes an Implicit Neural Representation (INR) framework to address these challenges by treating the image as a continuous mathematical function.
Methods
The core of the INR approach is representing the initial pressure distribution not as a grid of pixels, but as a continuous function $f(x, y) = \text{Intensity}$ parameterized by a Multi-Layer Perceptron (MLP).
Key technical highlights include:
- Self-Supervised Learning: The network does not require a database of prior scans. Instead, it is trained on the raw data of the current scan. It iteratively adjusts its internal weights until the image it represents produces ultrasound signals that match the actual recorded data.
- Hash Encoding: To capture fine details like tiny vessels efficiently, the authors utilize multi-resolution hash encoding to map spatial coordinates before they enter the MLP.
- Continuous Forward Model: Because the representation is continuous, the system can calculate expected ultrasound signals at any point in time without being limited by a fixed pixel grid, significantly reducing discretization errors.
Results
The study validated the INR approach using simulated vessel structures and physical phantoms containing black plastic spheres and tungsten wires. The INR method consistently outperformed both UBP and MB methods in both visual quality and quantitative metrics.
- Quantitative Gains: In experimental data, INR improved the Signal-to-Noise Ratio (SNR) by 6.83–19.42 dB and the Contrast-to-Noise Ratio (CNR) by 5.85–19.33 dB over competing methods.
- Structural Accuracy: In simulations, the INR method achieved a Structural Similarity Index (SSIM) of 0.99 with 256 projections, compared to 0.97 for MB and 0.37 for UBP.
- Artifact Suppression: Even with as few as 32 or 64 projections, the INR method resolved complex shapes (like "star" and "heart" phantoms) while UBP images were dominated by streaks. The continuity of the MLP acts as a natural regularizer, keeping the image smooth and suppressing background fluctuations.
Conclusion
The INR framework represents a significant step forward for sparse-view PACT. By moving away from pixel grids, it avoids the resolution-dependent ill-posedness of traditional methods and eliminates the need for large training datasets. While the current reconstruction time (approx. 8 minutes for 128 projections) is not yet real-time, it is already faster than some iterative MB implementations. Future work will focus on optimizing speed and extending the model to handle heterogeneous "Speed of Sound" (SOS) maps for improved performance in live tissue.
Summary updated