Multikernel positional embedding convolutional neural network for photoacoustic reconstruction with sparse data.

Area of Science:

• Biomedical Imaging
• Medical Technology
• Artificial Intelligence in Medicine

Background:

• Photoacoustic imaging (PAI) offers high contrast and resolution but suffers from low-quality reconstruction due to sparse data.
• Sparse spatial sampling and limited view detection hinder the widespread medical application of PAI.
• Deep learning has emerged as a promising solution for improving PAI reconstruction.

Purpose of the Study:

• To propose a novel deep learning architecture, DPM-UNet, for enhanced photoacoustic image reconstruction.
• To address the challenges of low-quality reconstruction in PAI caused by sparse data.
• To improve the imaging effect and reduce memory consumption in PAI.

Main Methods:

• Developed DPM-UNet, a U-Net based architecture incorporating position embedding, multi-kernel convolution blocks, and dilated dense blocks.
• Validated the DPM-UNet model using both simulated and in vivo photoacoustic imaging data.
• Compared the performance of DPM-UNet against other advanced reconstruction methods.

Main Results:

• Achieved high quantitative performance with a Structural Similarity Index Measure (SSIM) of 0.9824 and a Peak Signal-to-Noise Ratio (PSNR) of 33.2744 dB.
• Demonstrated superior image reconstruction quality compared to existing advanced methods.
• Showcased significant advantages in imaging effect and reduced memory consumption.

Conclusions:

• The proposed DPM-UNet architecture effectively enhances photoacoustic image reconstruction quality.
• DPM-UNet offers a significant advancement for medical applications of PAI, overcoming limitations of sparse data.
• The model provides a superior balance of imaging performance and computational efficiency.