Enhanced score-based priors-boosted generative model for resolution breakthrough in photoacoustic tomography

Photoacoustic tomography (PAT), a non-invasive biomedical imaging modality, integrates optical contrast with acoustic depth resolution. However, industry-standard ultrasonic transducers exhibit limited bandwidth, inherently degrading image resolution in conventional reconstruction. To address this, we propose a sinogram-domain signal spectral reconstruction method based on an enhanced score-based diffusion model (ESDM). The ESDM employs denoising score-matching to learn sinogram data distributions, augmented by an auxiliary network that captures cross-bandwidth dependency priors between full- and limited-bandwidth sinograms, thereby optimizing prior utilization during reconstruction. Evaluated against U-Net and Cycle-GAN using simulated and real-system experimental data, the proposed method demonstrates superior recovery of missing frequency components via back-projection. The resolution of the proposed method is about 1.60 times that of the limited-bandwidth baseline. This advancement significantly enhances PAT imaging under bandwidth constraints, broadening its clinical applicability.