Research on Deep Image Prior-Based Unmixing Method for Photoacoustic Imaging

Multispectral photoacoustic imaging can provide biochemical information closely associated with molecular metabolism, offering significant potential for early noninvasive disease diagnosis. In this paper, we present a deep learning-based unmixing technique for multispectral photoacoustic imaging. U-Net based on deep image prior is utilized to estimate the abundance of different components in biological tissues with the endmember prior and acquired multispectral photoacoustic datasets. To verify the feasibility of the method in photoacoustic imaging unmixing, numerical simulation experiments are conducted on two models for different biochemical components. The results show that the proposed method has a good effect on the unmixing of several different biomolecules, and realize the localization and abundance estimation of biomolecules. In addition, the unmixing results of photoacoustic images with different noises are still good,, which verify that the method is robust to Gaussian noise. Compared with traditional deep learning-based methods, the key advantages of this method lie in its ability to estimate abundance without requiring a large number of training datasets, coupled with its strong resistance to noise interference and competitive computational efficiency. It holds significant potential in the application of photoacoustic imaging unmixing and biological molecule component identification.