Simultaneous Denoising and Localization Network for Photoacoustic Target Localization.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Accurate localization of deep targets (vessels, needles, tumors) in photoacoustic (PA) images is crucial but challenging.
• Deep targets suffer from low signal-to-noise ratio (SNR) due to optical scattering in tissues.
• Existing methods struggle with noise and signal decay for deep targets.

Purpose of the Study:

• To develop a robust deep learning method for precise localization of deep targets in PA images.
• To enhance the signal-to-noise ratio (SNR) of photoacoustic radio-frequency (RF) data.
• To improve the accuracy of target localization in the presence of noise and optical scattering.

Main Methods:

• A novel deep neural network with a shared encoder and two parallel decoders was designed.
• One decoder estimates target coordinates, while the other enhances SNR and reconstructs clean RF data.
• Custom layers and regularizers were incorporated, and the network was trained on simulated datasets accounting for depth and scattering.

Main Results:

• The proposed deep learning method demonstrated significant robustness to noise in PA RF data.
• The network accurately localized targets in both simulated and experimental PA datasets.
• The method outperformed state-of-the-art techniques in target localization accuracy.

Conclusions:

• The developed deep learning approach effectively addresses the challenge of deep target localization in PA imaging.
• The method's noise robustness and SNR enhancement capabilities are critical for clinical applications.
• This work advances PA imaging by enabling more precise identification of subsurface structures.