Blind single image restoration for ring-array photoacoustic tomography system via physics-aided unsupervised learning

Ring-array photoacoustic tomography (PAT) has been successfully applied in noninvasive biomedical imaging. However, due to the influence of aperture shape and bandwidth of transducer, the reconstructed images usually suffer from spatially variant blur artifacts. Although deep learning methods can be employed for image restoration, they typically require large amounts of high-quality training data with ground truth to ensure good generalization, which is impractical for real-world ring-array PAT systems. In this study, we analyze and model the physics of blur formation in ring-array PAT system, integrate this with the unsupervised learning theory of deep prior (DP), and propose a blind single-image restoration framework. This framework consists of three generative networks designed based on the characteristics of the clear image and different types of point spread functions (PSFs). The parameters of the networks are optimized simultaneously to generate the optimal estimations of corresponding terms. Experimental results on both phantom and in vivo PAT images demonstrate that the proposed method is highly efficient and significantly improves the image resolution, outperforming state-of-the-art alternatives.