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

Area of Science:

• Biomedical Imaging
• Artificial Intelligence
• Medical Technology

Background:

• Photoacoustic imaging (PAI) offers high-resolution vascular visualization but faces limitations in spatial resolution, depth, and image quality due to noise and artifacts.
• Existing PAI techniques struggle with the trade-off between resolution and imaging depth, impacting diagnostic capabilities.

Purpose of the Study:

• To introduce GDSU-Net, a novel neural network for super-resolution (SR) reconstruction in PAI.
• To address the performance constraints of PAI by enhancing image resolution and quality.

Main Methods:

• GDSU-Net, based on the U-Net architecture, integrates group normalization, depthwise separable convolutions, SE blocks, and a pixelshuffle-based decoder.
• The model was fine-tuned for SR reconstruction of PAI data.

Main Results:

• GDSU-Net achieved a structural similarity index of 0.889 and a PSNR of 31.979 dB.
• The model significantly reduced RMSE to 0.032 and MAE to 0.025.
• Visual evaluations confirmed effective restoration of vascular details with anatomical fidelity.

Conclusions:

• GDSU-Net demonstrates superior performance in SR reconstruction for PAI.
• The developed model offers a computationally efficient solution for enhancing PAI resolution and image quality.
• GDSU-Net has the potential to improve diagnostic accuracy in applications relying on vascular imaging.