Robust denoising of low-fluence single-shot PAM using spatial attention-enhanced deep learning

Photoacoustic microscopy (PAM) enables label-free imaging of microvasculature with high optical absorption contrast and distinctive spatial resolution. However, time-resolved applications, specifically in vivo monitoring of microcirculation and high-speed volumetric scanning, often suffer from low signal-to-noise ratio (SNR) due to limited laser fluence and single-acquisition noise. In this study, we propose a volumetric denoising framework based on a spatial attention-enhanced 3D U-Net. By incorporating a spatial attention mechanism into 3D patch-based learning, the network effectively exploits volumetric spatial correlations to distinguish vascular features from background noise. To address data scarcity, we employ a transfer learning strategy: pre-training on leaf vein phantoms, followed by fine-tuning on in vivo mouse ear vasculature. Experimental results demonstrate that, compared with single-shot images, our method significantly improves the peak signal-to-noise ratio (PSNR) from 18.93 dB to 29.76 dB and the structural similarity index (SSIM) from 0.3943 to 0.9008, with an SNR enhancement of over 13 dB. This approach enables high-fidelity PAM imaging suitable for low-fluence or rapid acquisition scenarios.