Originally published in:
Radiology 2019;290(3):649–656
https://doi.org/10.1148/radiol.2018180940
Erratum in:
Online only: DOI:10.1148/radiol.2020202527
This erratum corrects the network schematic in Figure 1. The network structure in Figure 1 should be as follows: The encoder portion consists of three sets of three 3 × 3 convolution (conv)-batch normalization (BN)-rectified linear unit activation (ReLU) operations, with 32, 32, and 64 tensors for the convolutions in each set respectively; 2 × 2 max-pooling is performed on the output of each set and fed into the next set. The center connection consists of one set of three 3 × 3 conv (64 tensors)-BN-ReLU operations; the result is added with the input of the center connection (residual connection) and passed on to the decoder portion. The decoder portion consists of three sets of three 3 × 3 conv-BN-ReLU operations, with 64, 32, and 32 tensors for the convolutions in each set respectively. The inputs of each set are a concatenation of the output of the previous set after 2 × 2 up-sampling and the output of its corresponding encoder set. A 1 ×1 convolution with hyperbolic tangent activation is performed on the output of the final decoder set and added to the original low-dose PET input to obtain the output image.