Table 1.
Example direct AI reconstruction methods
| Name | Architecture [total parameters] | Input | Target | Loss function and optimiser (epochs) validation | Number of samples (training, validation, test) |
|---|---|---|---|---|---|
| AUTOMAP Zhu et al. 2018 10 |
Two fully connected layers and CNN [∼800M] | 2D noisy sinograms | T1w brain images (128 × 128) |
MSE with L1-norm penalty on network weights in final hidden layer RMSProp (100) Validation not used |
50,000 n/a 1 |
| DeepPET Häggström et al. 2019 11 |
CED [>60M] |
2D noisy sinograms (269 × 288) | Ground-truth PET images (128 × 128) |
MSE Adam (150) Validation used |
203,305 (70%) 43,499 (15%) 44,256 (15%) |
| DPIR-Net Hu et al. 2020 39 |
CED [>60M] Discriminator [>3.5M] |
2D noisy sinograms (269 × 288) | Ground-truth PET images (128 × 128) |
Wasserstein GAN + VGG + MSE Adam (100) Validation not used |
37,872 (80%) n/a 9468 (20%) |
| CED extended to SSRB sinograms from large FOV PET Ma et al. 2022 40 |
CED [∼64M] |
2D noisy sinogram (269 × 288) | Reconstructed PET images using OSEM + PSF TOF from list-mode data (128 × 128) |
MSE + SSIM+ VGG Adam (300) Validation used |
35,940 (76%) 5590 (12%) 5590 (12%) |
| FastPET Whiteley et al. 2021 48 |
U-Net [∼20M] |
Noisy histo-image slices + attenuation map slices (2 × 440 × 440 × 96) |
Image slices reconstructed using OSEM + PSF (440 × 440 × 96) |
MAE + MS-SSIM Adam (500) Validation used |
20,297 slices (74%) 1767 slices (6%) 5208 slices (20%) |
CNN refers to a convolutional neural network, CED refers to a convolutional encoder-decoder. VGG in this table refers to perceptual loss based on a VGG network.
3D, three-dimensional; FBP, filtered backprojection; MLEM, maximum likelihood–expectation maximisation; OSEM, ordered subsets expectation maximisation; PSF, point spread function.