A Deep Learning Approach to Photoacoustic Wavefront Localization in Deep-Tissue Medium

Area of Science:

• Biomedical Optics
• Medical Imaging
• Machine Learning

Background:

• Optical photons scatter significantly in biological tissues, hindering deep imaging.
• Photoacoustic imaging (PAI) offers high-resolution, label-free visualization of vasculature but is limited by depth-dependent optical attenuation.
• Accurate localization of photoacoustic signal origins is crucial for deep-tissue applications.

Purpose of the Study:

• To develop a robust method for localizing photoacoustic signal origins in optically scattering deep tissues.
• To overcome the limitations of depth-dependent optical attenuation in photoacoustic imaging.
• To enhance the visibility of deep-tissue vasculature and other light-absorbing targets.

Main Methods:

• An encoder-decoder convolutional neural network (CNN) architecture with custom modules was designed and trained.
• The CNN was trained on simulated photoacoustic signals from numerous blood vessel targets under scattering and noise conditions.
• An ablation study validated the contribution of each network module to localization accuracy.

Main Results:

• The CNN achieved high localization accuracy for photoacoustic targets in scattering media.
• Mean localization error was <30 microns for targets <40 mm deep and 1.06 mm for targets 40-60 mm deep.
• Experimental validation confirmed the network's performance under various scattering conditions.

Conclusions:

• The proposed deep learning approach effectively breaks through the optical diffusion limit in photoacoustic imaging.
• This method significantly improves the localization of deep-tissue photoacoustic sources.
• Potential applications include optical wavefront shaping, melanoma cell detection, and enhanced vascular surgery.