A machine learning method to identify and remove reflection artifacts in photoacoustic channel data

Photoacoustic imaging is often used to visualize point-like targets, including circular cross sections of small cylindrical implants like brachytherapy seeds as well as circular cross sections of metal needles. When imaging these pointlike targets in the presence of highly echogenic structures, the resulting image will suffer due to reflection artifacts which appear as true signals in the traditional beamformed image. We propose to use machine learning methods to identify these types of noise artifacts for removal. A deep convolutional neural network was trained to locate and classify source and reflection artifacts in photoacoustic channel data simulated in k-Wave. Simulated channel data contained one source and one artifact with varying target locations, medium sound speeds, and -3dB channel noise. In testing 3,998 simulated images, we achieved a 99.1% and 98.8% success rate in classifying sources and artifacts, respectively, while obtaining a misclassification rate below 3.1%, where a misclassification was defined as a source or artifact detected as an artifact or source, respectively. The network, which was only trained on simulated data, was then transferred to experimental data with 100% source classification accuracy and 0.40 mm mean source location accuracy. These results are promising as they show that a network trained with only simulated data can distinguish experimental sources and artifacts in photoacoustic channel data and display this information in a novel artifact-free image format.