Deep Learning-Based Super-Resolution for Vessel Enhancement in Photoacoustic Microscopy Imaging.

Abstract

Photoacoustic imaging (PAI) is an advanced imaging technique for high-resolution (HR), non-invasive visualization of vascular networks, offering distinct advantages in functional and structural imaging. However, its performance is often constrained by trade-offs between spatial resolution and imaging depth, as well as noise and artifacts caused by system limitations and tissue properties. Here, we introduce GDSU-Net, a fine-tuned neural network designed for super-resolution (SR) reconstruction of PAI. GDSU-Net builds on the U-Net architecture and incorporates four key components: group normalization, depthwise separable convolutions, squeeze-and-excitation (SE) blocks, and a pixelshuffle-based decoder. Experimental results demonstrate that GDSU-Net achieves a structural similarity index of 0.889 and a peak signal-to-noise ratio (PSNR) of 31.979 dB, while reducing the root mean square error (RMSE) to 0.032 and the mean absolute error (MAE) to 0.025. Visual evaluations confirm its effectiveness in restoring vascular details with preserved anatomical fidelity. These findings highlight GDSU-Net as a computationally efficient solution for super-resolution enhancement in PAI.