High Sensitivity Photoacoustic Imaging by Learning From Noisy Data.

Area of Science:

• Biomedical Imaging
• Deep Learning
• Computational Photography

Background:

• Photoacoustic imaging (PAI) offers high-resolution, non-invasive detection of biological tissues.
• Image quality in PAI is often limited by low signal-to-noise ratio (SNR) due to weak signals, low chromophore concentration, and noise.
• Existing hardware and computational solutions for SNR improvement are often costly or lack generalizability.

Purpose of the Study:

• To develop a self-supervised deep learning method for enhancing PAI SNR.
• To provide a cost-effective and broadly applicable solution for improving PAI image quality.
• To demonstrate the method's effectiveness in visualizing deep tissue structures and tumors.

Main Methods:

• A self-supervised deep learning algorithm was developed to increase SNR in PAI data.
• The method was trained exclusively on noisy photoacoustic images, eliminating the need for ground truth data.
• The approach was validated on both microscopic and computed tomographic PAI datasets from various systems.

Main Results:

• The method significantly improved SNR in photoacoustic images, with increases up to 12-fold.
• Deep vascular details, previously obscured by noise, became clearly visible.
• Imaging depth was effectively doubled, and high-contrast imaging of deep tumors was achieved.

Conclusions:

• The presented self-supervised deep learning method offers an accessible and effective way to enhance PAI SNR.
• This technique can overcome limitations of current methods, enabling better visualization in preclinical and clinical settings.
• The approach holds promise for broader application across diverse PAI systems and research areas.