PAT-Mamba: Adaptive all-in-one image restoration for photoacoustic tomography via state space model.

Photoacoustic tomography (PAT) combines optical contrast with ultrasonic penetration, yet hardware limitations and workflow constraints in practical systems inevitably introduce multiple degradation pathways including channel under-sampling, angular under-sampling, bandwidth limitations, and additive noise that severely deteriorate reconstructed image quality. Existing learning-based solutions typically specialize in a single degradation and therefore demand multiple models in practice. We present PAT-Mamba, a novel all-in-one blind restoration framework that simultaneously addresses multiple prevalent degradation types without prior knowledge of the acquisition conditions. PAT-Mamba incorporates two complementary innovations: TransMamba blocks that replace self-attention with selective state space models to capture long-range dependencies at linear complexity, enabling efficient global context propagation for artifact suppression and structural detail recovery, and Adaptive Frequency Learning Blocks (AFLB) that analyze degradation-specific frequency cues, generate learnable frequency masks, and refine frequency subbands through cross-attention and dual-branch gating. Experiments on three representative datasets demonstrate that PAT-Mamba achieves competitive or improved performance compared to state-of-the-art baselines in the majority of degradation scenarios, attaining an average PSNR of 37.04 dB on the Embryo simulation dataset (+0.91 dB over the runner-up), 30.41 dB on the in vivo Mice dataset (+0.53 dB), and 30.50 dB on the clinical Artery dataset. These results demonstrate that PAT-Mamba eliminates the need to deploy multiple specialized networks for different degradation types, while its blind restoration capability removes the requirement for prior degradation diagnosis in clinical workflows, providing a practical and efficient restoration solution for real-world PAT systems. Code is available at https://github.com/FDUlixiaoyao/PAT-Mamba.