U-Net enhanced real-time LED-based photoacoustic imaging
Avijit Paul1, Srivalleesha Mallidi1
1Department of Biomedical Engineering, Tufts University, Medford, Massachusetts, USA.
Journal of Biophotonics
|April 16, 2024
Summary
This study introduces a U-Net deep learning framework to improve photoacoustic (PA) imaging quality. The method enhances signal-to-noise ratio (SNR) and contrast, enabling faster imaging with low-energy light sources.
Area of Science:
- Medical Imaging
- Biomedical Engineering
- Artificial Intelligence
Background:
- Photoacoustic (PA) imaging combines optical contrast and spatial resolution.
- Light-emitting diodes (LEDs) are emerging as cost-effective PA optical sources.
- Low light fluence from LEDs necessitates frame averaging, reducing acquisition speed.
Purpose of the Study:
- To develop a deep learning framework for enhancing PA image quality.
- To improve signal-to-noise ratio (SNR) and contrast in PA images.
- To enable faster PA imaging using low-energy light sources.
Main Methods:
- A U-Net deep learning architecture was implemented.
- The framework processes PA images acquired with low frame averaging.
- The model was trained and validated on in vitro phantoms and in vivo models.
Main Results:
- The U-Net framework achieved a four-fold increase in SNR for both in vitro (4.39 ± 2.55) and in vivo (4.27 ± 0.87) data.
- The network demonstrated noise invariance.
- Potential downsides include blurry outcomes and failure to reduce salt-and-pepper noise.
Conclusions:
- The developed U-Net framework effectively enhances PA image SNR and contrast.
- This deep learning approach facilitates real-time image enhancement for low-cost PA systems.
- The method supports the clinical translation of LED-based PA imaging.
