Quantitative Photoacoustic Blood Oxygenation Imaging Using Deep Residual And Recurrent Neural Network

摘要

Diffusive light scattering in biological tissue causes a heterogeneous and variable spectrum of light fluence, which is a significant challenge to predict due to unknown optical properties of tissues. Then it is difficult to accurately and quantitatively image blood oxygen saturation (sO₂) in photoacoustic imaging because of unknown fluence distribution. To tackle this problem, we develop a deep residual, and recurrent neural network, i.e., DR2U-net for the quantitative estimation of blood oxygenation of photoacoustic imaging. The fine-tuned DR2U-net we developed can extract fluence distribution information from optical absorption images only using two wavelengths of light in Monte Carlo simulation, and afterward generate a quantitative image of sO₂. The measurement results of sO₂ show a very high accuracy by testing in simulated biological tissue, and its error is as low as 1.27% compared with conventional linear mixing method (48.76%). Besides, our model is high-speed, about 18. 4ms is achieved for every quantitative sO2 image.