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

Low-cost optoacoustic imaging based on light-emitting diodes (LEDs) offers an affordable alternative to traditional laser-based systems, potentially broadening the reach of this technology into resource-limited settings. However, LEDs are only able to excite very weak optoacoustic responses, which leads to prominent noise artifacts in the reconstructed images.

We aim to mitigate noise-related artifacts in LED-based optoacoustic tomography and thereby enhance the image quality and usability of these low-cost systems.

We propose a squeeze-and-excitation U-Net-based model (SE-UNet) for noise artifact reduction. The network incorporates a VGG19 convolutional neural network mid-layer feature extractor as a loss evaluation module. It is trained on noisy data paired with high-quality reference images generated using a conventional solid-state pulsed laser source.

Our model achieves consistent improvements on no-reference image-quality metrics (NIQE and BRISQUE) and in the contrast-to-noise ratio, effectively reducing noise artifacts while preserving image structure and details. In addition, it exhibits a rapid processing time of $\sim 3\mathrm{ms}$ per $480\times 480\text{pixel}$ image on a GTX 2070 GPU, utilizing 627 MB of memory.

These results highlight the potential of the proposed SE-UNet model for optimizing the performance of LED-based optoacoustic imaging systems, offering both high efficiency and improved image quality.