Deep Neural Networks to Remove Photoacoustic Reflection Artifacts in Ex Vivo and in Vivo Tissue

Deep neural networks trained with simulated data are capable of distinguishing sources from reflection artifacts in photoacoustic data. Our group recently introduced this concept with simulated and experimental waterbath and phantom data. In this novel approach, channel data is used as an input to learn the spatial impulse response of pressure waves from point-like sources and differentiate true sources from reflection artifacts. We hypothesize that this is possible based on learned differences in the unique shape-to-depth relationship of point sources. The work presented in this paper builds on previous demonstrations to investigate the feasibility of this approach when applied to ex vivo tissue and in vivo data from a pig catheterization procedure. Three networks were trained with k-Wave simulated data: two residual deep network architectures (i.e., Resnet-50 and Resnet-101) and the previously implemented VGG16 deep network architecture, which does not use residual learning. These networks classified sources correctly in over 82% of images in the ex vivo chicken breast, liver, and steak datasets and over 64% of images in the dataset acquired from an ex vivo chicken thigh containing bone. When applied to in vivo data, the Resnet-50 and Resnet-101 architectures classified 83.3% and 88.8% of sources correctly, respectively, while the VGG16 architecture performed more poorly, classifying 14.5% of sources correctly. In addition, the residual network architectures had <2% misclassification rate, whereas the VGG16 architecture had a maximum 23.53% misclassification rate for all datasets. These results indicate that residual networks architectures are better suited to in vivo source detection and artifact elimination using our approach.