A generative adversarial network with multi-scale structural features for sparse-view photoacoustic tomography reconstruction

Photoacoustic tomography combines high optical absorption contrast with deep ultrasonic penetration, providing powerful capabilities for high resolution biomedical imaging. It often demands wide field and large data acquisition, posing significant challenges to imaging speed. Sparse sampling effectively reduces data volume and accelerates imaging, but the resulting artifacts and detail loss limit its applications. Therefore, we construct a sparse-view photoacoustic tomography system and propose generative adversarial network with multiscale structural features, to achieve high speed and quality three-dimensional imaging. The generator incorporates pyramid squeeze attention block to extract multiscale structural information, while skip connection with a channel attention module in the discriminator enhances organ structural consistency. In addition, dual gradient regularized adversarial loss is designed to improve stability in detail enhancement and artifact suppression. Experiments with 128 and 64 views sampling verify that the proposed system achieves superior artifact suppression and detail recovery, preserving structural features consistent with full-view reconstruction. Furthermore, compared with the strongest baseline diffusion model and MambaIR, MSF-GAN improved the peak signal-to-noise ratio and structural similarity index by 1.128% and 1.176% under 128 views sampling, and by 2.442% and 0.536% under 64 views. These demonstrate that the proposed system and method effectively leverage photoacoustic structural information to improve fidelity and contrast. This work achieves a desirable balance between high speed and quality reconstruction, providing a promising pathway for translating photoacoustic tomography into practical clinical applications.