Figure 10.

Image reconstructions assisted by machine learning. (a–e) Deep learning PAT with sparse data.¹¹⁵ (a) The U-Net network architecture, consisting of contracting (downsampling) and expansive (upsampling) paths, which is used for the image reconstruction with sparse data. (b) Artefactual reconstructed image with undersampled (128 projections) data, showing the reconstruction artifacts due to the sparse data. (c) Zoom-in images of the yellow and green boxed regions in (b). (d) Artifact-free counterpart of (b), obtained with the trained network. (e) Zoom-in images of the yellow and green boxed regions in (d). (f, g) Hybrid neural network for limited-view PACT.¹¹⁸(f) The global architecture of Y-Net. Two encoders extract different input features, which concatenate into the decoder. Both encoders have skip connections with the decoder. (g) Comparison of reconstructed images. Top left, ground truth; top right, image reconstructed using the universal back-projection method; bottom left, image reconstructed using the time-reversal method; bottom right, image reconstructed using the trained Y-net.