Multi-wavelength graph convolutional network for high-performance sparse multispectral optoacoustic tomography

The rapid advancement of multispectral optoacoustic tomography (MSOT) has developed for label-free biomedical imaging by providing anatomical and functional visualization through multi-wavelength laser excitation and ultrasound detection. This technique offers high spatial resolution and deep-tissue imaging capabilities for biological applications. However, the substantial hardware cost and computational demand for high-quality in vivo imaging hinder its extensive development. To overcome these limitations, we propose a multi-wavelength graph convolutional network for sparse MSOT. Our approach solves the ill-conditioned sparse reconstruction problem through a graph learning framework integrated with a multi-wavelength sparse sampling strategy, which can model and leverage the intrinsic correlations in artifact distributions across diverse sparse transducer configurations. Comprehensive in vivo mouse experiments demonstrate that the proposed method provides a flexible and practical solution for high-performance sparse MSOT imaging under sparse conditions (16 transducer elements with the reconstruction SSIM of 0.92 ± 0.01 and PSNR of 27.74 ± 1.27).