End-to-end deep neural network for optical inversion in quantitative photoacoustic imaging

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Quantitative photoacoustic imaging requires accurate methods for analyzing tissue composition.
• Deep neural networks offer potential for complex image reconstruction tasks.

Purpose of the Study:

• To develop and evaluate an end-to-end deep neural network, ResU-net, for quantitative photoacoustic imaging.
• To assess the accuracy and robustness of ResU-net in estimating oxygen saturation (sO₂) and chromophore concentration.

Main Methods:

• An end-to-end deep neural network, ResU-net, was developed utilizing a residual learning framework.
• The network processed multispectral initial pressure images to infer quantitative chromophore concentration and sO₂ images.

Main Results:

• ResU-net demonstrated accurate and robust estimations of sO₂ and indocyanine green concentration.
• The model showed resilience against variations in optical properties and object geometry.
• An exceptionally fast reconstruction time of 22 milliseconds was achieved.

Conclusions:

• ResU-net provides an efficient and accurate solution for quantitative photoacoustic imaging.
• The deep learning approach facilitates improved analysis of tissue oxygenation and chromophore distribution.