Efficient denoising in LED-based optoacoustic tomography with squeeze-and-excitation deep convolutional networks

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Low-cost optoacoustic imaging utilizes light-emitting diodes (LEDs) as an affordable alternative to laser-based systems.
• LEDs generate weak optoacoustic signals, resulting in significant noise artifacts in reconstructed images.
• These artifacts limit the usability and diagnostic potential of LED-based optoacoustic systems, especially in resource-limited environments.

Purpose of the Study:

• To mitigate noise artifacts in LED-based optoacoustic tomography.
• To enhance the image quality and practical utility of low-cost optoacoustic imaging systems.
• To enable broader adoption of optoacoustic technology in settings with limited resources.

Main Methods:

• Development of a squeeze-and-excitation U-Net (SE-UNet) deep learning model for noise reduction.
• Integration of a VGG19 convolutional neural network as a feature extractor for loss evaluation.
• Training the SE-UNet model using paired noisy LED-excited data and high-quality laser-excited reference images.

Main Results:

• The SE-UNet model consistently improved no-reference image quality metrics (NIQE, BRISQUE) and contrast-to-noise ratio.
• Noise artifacts were effectively reduced while preserving crucial image structures and fine details.
• Rapid image processing achieved with a processing time of approximately 3 ms per 480x480 pixel image on a GTX 2070 GPU.

Conclusions:

• The proposed SE-UNet model demonstrates significant potential for optimizing LED-based optoacoustic imaging.
• The approach offers a viable solution for enhancing image quality and efficiency in low-cost systems.
• This advancement could facilitate wider clinical and research applications of optoacoustic imaging, particularly in underserved areas.