Implicit neural representation for sparse-view photoacoustic computed tomography

High-quality imaging in photoacoustic computed tomography (PACT) requires a high-channel count system for dense spatial sampling 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, the explored methods use discrete representation to formulate the reconstruction problem, which is ill-conditioned and is thus prone to measurement errors and discretization errors. Meanwhile, the extent of the ill-condition of the reconstruction 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 to address these issues. 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 then determined by a training process in a self-supervised manner, by minimizing the errors between the measured and model-predicted PA signals. After training, PA images can be mapped by feeding the coordinates to the network. Simulation and phantom experiments showed that the INR method outperforms the universal back-projection and MB methods in preserving image fidelity and artifacts suppression for the same acquisition condition. In the experimental data, the INR method improved signal-to-noise-ratio (generalized contrast-to-noise-ratio) by 1.1-24.0 dB (0.037-0.716), compared to the other methods. These results clearly demonstrated the value of INR for high-quality PACT image reconstruction with sparse data and its potential in reducing the complexity of PACT systems.