Photoacoustic (PA) imaging combines optical spectroscopic contrast with deep tissue penetration, offering valuable functional, molecular, and structural information about tissue. However, a long-standing challenge with PA imaging has been that the quantification accuracy of tissue chromophores concentrations remains limited due to the spectral colouring effect. Monte Carlo (MC) simulation is regarded as the gold standard to model light transportation in tissue but can be computationally demanding, thus not suitable for real-time applications. We propose a time-efficiency solution using conditional generative adversarial networks (cGANs) to generate light fluence distributions within tissue towards real-time spectral decolouring in PA imaging. The networks were trained to predict light fluence distribution from realistic tissue anatomy and optical properties using MC simulation as ground truth. We achieved high-quality light fluence synthesis, with a peak signal-to-noise ratio of 31.9 dB using in vivo segmentation. We also demonstrated the validity of spectral decolouring for PA quantification, with an error of absorption efficient estimation around 0.05 using numerical phantoms. Thus, this approach holds promise for enhancing the quantification performance of PA imaging in real-time.
|