Table 2.

Unrolled methods for AI in PET image reconstruction.

Name	Architecture [total parameters]	Input	Target	Loss function and optimiser (epochs) validation training method	Number of samples (training, validation, test)
EM-Net Gong et al. 2019 54	10 modules U-Net shared [∼2M]	Previous output/iteration	3D high count reconstruction (128 × 128 × 46)	MSE Adam Validation not used Gradient truncation	16 n/a 1
MAPEM-Net Gong et al. 2019 61	8 modules U-Net not shared [∼16M]	Current output from a block	3D high count reconstruction (128 × 128 × 105)	MSE [Details not specified] Validation not used End-to-end	18 n/a 1
FBSEM-Net Mehranian and Reader 2020 52	10 modules CNN shared [∼77k]	Previous output / iteration	3D high count reconstruction— cropped (114 × 114 × 128)	MSE Adam (200) Validation used Gradient truncation	45 5 5
Iterative Neural Network Lim et al. 2020 56	10 modules CNN not-shared [∼40k]	Current output from a block	3D true activity image (200 × 200 × 112)	MSE Adam (500) Validation not used Sequential training	4 n/a 1
TransEM Hu and Liu 2022 62	10 modules Swin Transformer [details not specified]	Previous output / iteration	2D high count reconstruction	MSE Adam Validation used Gradient truncation	510 30 60

AI, artificial intelligence; CNN, convolutional neural network; 2D, two-dimensional; 3D, three-dimensional; PET, positron emission tomography.