Deep learning reconstruction algorithm based on sparse photoacoustic tomography system

Limited by physical architecture and cost, devices that use limited viewing angles and limited transducers to detect imaging targets are common in photoacoustic tomography (PAT), which can lead to the loss of target information and make the images reconstructed by traditional algorithms contain a lot of Artifact. This paper designs a new neural network architecture, as well as a suitable algorithm evaluation system and training strategy, and achieves high-quality image reconstruction on the actual sparse system. The signal conversion network is used to extract the characteristics of the original sound pressure data and complete the conversion from the signal domain to the image domain. In order to obtain a better reconstruction effect, the signal feature map is combined with a low-quality linear reconstruction image. The improved U-net network is used to further process the combined image. Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) is used to comprehensively evaluate the multi-evaluation index scheme to retain the best network model. This method is verified by simulation and actual experiment, and it has been proved to be able to reconstruct high-quality images from sparsely measured analog data.