Self-supervised Spatial Interpolation for 3D Photoacoustic Imaging Using Deep Learning

The three-dimensional(3D) photoacoustic (PA) imaging (PAI) system has shown its advantages in volumetric imaging capabilities. However, due to limited resources, the 3D PAI system generally has sparse sensors that cause degradation in image reconstruction, demanding effective interpolation algorithms for versatile 3D PAI systems with different configurations. However, existing interpolation methods generally rely on regular sensor distribution, making them inapplicable to be used in 3D PAI systems with complex distributions. Here, we propose a self-supervised interpolation algorithm for the 3D PAI system based on a graph Neural Network (GNN). The structure of GNN is constructed by connecting each sensor node, and the network learns to map the spatial positions of these sensors to their time-domain PA signals. Our method only requires the data that need to be interpolated as input, and predicts time-domain PA signals for sensors at new positions after training, no need for extra large amount of pre-training PA data. In addition, the interpolation process is fast and is not affected by the distribution of sensors, which is important for its actual implementation. In phantom and in vivo animal studies, it demonstrated superior performance in both signal and image domains compared to traditional methods.