Self-supervised spatial interpolation for 3D photoacoustic imaging using deep learning

Three-dimensional (3D) photoacoustic imaging (PAI) systems offer significant advantages for volumetric imaging. However, limited hardware resources often lead to sparse sensor configurations, which degrade image reconstruction quality and create a need for effective interpolation algorithms adaptable to versatile 3D PAI systems with different sensor arrangements. Existing interpolation methods generally rely on regular sensor distributions, which limits their applicability in systems with complex geometries. In this work, we propose a self-supervised interpolation algorithm for 3D PAI based on a graph neural network (GNN). The GNN is constructed by treating sensors as nodes and connecting them according to their spatial relationships, enabling the network to learn the mapping from sensor positions to time-domain photoacoustic signals. The proposed method requires only the data to be interpolated as input and predicts time-domain PA signals at new sensor positions after training, without the need for a large amount of additional pretraining data. Moreover, the interpolation process is fast and insensitive to sensor distribution, which is important for practical implementation. Phantom and in vivo animal experiments demonstrate that the proposed method achieves superior performance in both signal and image domains compared with traditional methods.