DPAT-Net: A sequential dual-domain network with axial attention for limited-view photoacoustic image reconstruction

Photoacoustic imaging (PAI) under limited-view conditions often suffers from severe reconstruction artifacts due to incomplete data, compromising image interpretation and quantitative analysis. In this paper, we propose DPAT-Net, an end-to-end reconstruction network that combines known acquisition geometry with learned measurement- and image-domain priors. The network employs a "dual-domain processing" strategy: in the data domain, a hybrid sinogram enhancement module with bottleneck axial attention processes limited-view sensor-time data by modeling long-range sensor-axis dependencies; subsequently, a fixed-geometry Differentiable Back-Projection (DBP) layer maps both the enhanced sensor-time data and the raw limited-view measurements into the image domain, while allowing image-domain gradients to reach the upstream enhancement module; finally, an Image Refinement Network fuses the two preliminary images to produce the final reconstruction. Validation on a vascular simulation dataset, an in vivo mouse-brain dataset, and the public PATATO preclinical phantom dataset demonstrates higher PSNR and SSIM than the evaluated baseline methods under limited-view sampling. Notably, on the in vivo dataset, the proposed method achieves the highest PSNR and SSIM relative to the full-view reference reconstruction among the evaluated methods using only one-quarter of the measurement data (32 channels), with an inference time of 9.97 ms per image on an RTX 4090 GPU.