O-Net: A Convolutional Neural Network for Quantitative Photoacoustic Image Segmentation and Oximetry

TL;DR

This paper introduces a deep neural network that rapidly and accurately estimates blood oxygenation and segments vessels in photoacoustic images, enabling real-time quantitative mapping in biomedical applications.

Contribution

A novel deep neural network architecture that simultaneously performs photoacoustic image segmentation and oxygen saturation estimation with high speed and accuracy.

Findings

Median error of 5.1% in SO2 estimation

Segmentation accuracy better than 95%

Estimation time less than 50 milliseconds

Abstract

Estimation of blood oxygenation with spectroscopic photoacoustic imaging is a promising tool for several biomedical applications. For this method to be quantitative, it relies on an accurate method of the light fluence in the tissue. This is difficult deep in heterogeneous tissue, where different wavelengths of light can experience significantly different attenuation. In this work, we developed a new deep neural network to simultaneously estimate the oxygen saturation in blood vessels and segment the vessels from the surrounding background tissue. The network was trained on estimated initial pressure distributions from three-dimensional Monte Carlo simulations of light transport in breast tissue. The network estimated vascular SO2 in less than 50 ms with as little as 5.1% median error and better than 95% segmentation accuracy. Overall, these results show that the blood oxygenation can be quantitatively mapped in real-time with high accuracy.