Integrating visual state space models with frequency selection for artifact reduction in photoacoustic tomography

Photoacoustic tomography (PAT) with sparse angular sampling suffers from structured artifacts and loss of fine details. We propose the Visual Mamba and Frequency Selection Network (VMFSNet), an image-domain restoration framework that combines visual state-space modeling for long-range spatial dependencies with content-adaptive frequency selection. The frequency-selection module dynamically generates low-pass and residual high-pass features, enabling input-conditioned frequency modulation without predefined transform bases. A compound loss further constrains spatial intensity, Fourier-domain consistency, and structural similarity. VMFSNet was evaluated on simulated vessel, phantom, and in vivo mouse datasets at 32, 64, and 128 projections. Experimental results demonstrate strong overall performance across different sparse-view conditions, while ablation analyses support the complementary contributions of state-space modeling and adaptive frequency selection. These results demonstrate the potential of VMFSNet for efficient sparse-view PAT artifact reduction.