OGRR-Net: observation-guided residual refinement learning for photoacoustic imaging

Photoacoustic imaging provides optical-absorption contrast through ultrasonic detection. However, image reconstruction remains affected by acquisition geometry, detector response, acoustic noise, and the attenuation of high-frequency components during analytic inversion. Although deep post-processing methods have shown effectiveness in artifact reduction, purely image-domain restoration lacks a direct mechanism for incorporating complementary information contained in the raw measurement data. This work proposes OGRR-Net, an observation-guided residual refinement network for photoacoustic imaging. The network takes an initial reconstruction as the spatial reference and encodes the corresponding raw sinogram into an observation-guidance map, which conditions a gated residual correction pathway for image refinement. Evaluated on an in vivo mouse dataset and the OADAT optoacoustic dataset against representative image-domain and hybrid reconstruction methods, the proposed observation-guided strategy improves reconstruction fidelity, reduces artifacts, and enhances the preservation of structural details.