Table 4.
List of original deep learning methods evaluated on clinical PET data
| Study | Method | Training / validation / testing data† (Tracer) | Region | Reported error (%) | |
|---|---|---|---|---|---|
| Proposed method | Vendor method | ||||
| Ribeiro et al. [233] | Generation of template-based μ-maps [168] from UTE images using a three-layer network |
4/N.A/N.A (2-[18F]FDG) |
Brain | 3.4 | N/A |
| Bradshaw et al. [269] | Generation of 4-class probability maps from T1 and LAVA Flex images using the 3D DeepMedic network |
12/6/N.A (2-[18F]FDG) |
Pelvis | − 1.0 ± 1.3 (lesions) | Dixon: 0.0 ± 6.4 |
| Gong et al. [214] | Generation of pseudo-CTs from Dixon and ZTE images using a 2D U-Net with group convolutional modules |
32/8/N.A (2-[18F]FDG) |
Brain | ~ 2.0 ± 0.5 ^ |
*DixonSB: ~ 5.5 ± 1.5 ZTE: ~ 4 ± 1.3 |
| Jang et al. [231] | Generation of 3-class from rapidly acquired UTE images using a 2D VGG16 (encoder) and SegNet (decoder) |
30 T1 + 6 UTE (transfer learning)/ N.A/ 8 (2-[18F]FDG) |
Brain | 0.2 ± 1 | N/A |
| Liu et al. [229] | Generation of 3-class pseudo-CTs from AC PET 2-[18F]FDG images using a 2D VGG16 |
100/28/N.A (2-[18F]FDG) |
Brain | −0.6 ± 2.0 □ | N/A |
| Liu et al. [114] | Same as [229] but using a 13-layer VGG16 with T1 images as input |
100/28/N.A (2-[18F]FDG) |
Brain | −0.7 ± 1.1 □ | Dixon: −5.8 ± 3.1 |
| Arabi et al. [38] | Generation of 4-class μ-maps from T1 images using the 3D adversarial semantic structure |
15/15/N.A (2-[18F]FDG) |
Brain | 3.2 ± 3.4 (whole head) □ | N/A |
| Blanc-Durand et al. [75] | Generation of pseudo-CTs from ZTE images using a 3D U-Net algorithm |
50/43/N.A (2-[18F]FDG) |
Brain | −0.2 ± 5.6 □ | ZTE:2.5 |
| Dong et al. [247] | Generation of pseudo-CTs from NAC PET images using a 3D cycleGAN |
80/39/N.A (2-[18F]FDG) |
Whole body | − 1.1 ± 3.9 (Brain), 10.7 ± 7.7 (Lung), 0.7 ± 8.4 (Heart) | N/A |
| Hwang et al. [253] | Generation of CT-based μ-maps from MLAA images using a 2D U-Net |
60/20/20 (2-[18F]FDG) |
Whole body |
− 2.2 ± 1.78 (bone lesions) 1.3 ± 3.3 (soft-tissue lesions) |
Dixon: −9.4 ± 5.2 Dixon: −2.9 ± 1.2 |
| Ladefoged et al. [212] | Generation of pseudo-CTs for paediatric data from UTE, echo images and the R2* map using a 3D U-Net | 60/19/28 ([18F]FET) | Brain | −0.1 | N/A |
| Shiri et al. [220] | Generation of AC from NAC PET 2-[18F]FDG images using a 2D U-Net algorithm | 91/20/18 | Brain | − 0.1 ± 2.14 | N/A |
| Spuhler et al. [226]** | Generation of pseudo-transmission data from a spoiled gradient recalled acquisition using a 2D U-Net | 66/11/11 ([11C]WAY-100635) + 10 ([11C]DASB) | Brain |
− 0.5 ± 1.7 ([11C]WAY-100635) − 1.5 ± 0.7 ([11C]DASB) □ |
N/A |
| Torrado-Carvajal et al. [241] | Generation of pseudo-CTs from Dixon images as input to a 2D U-Net algorithm |
15/4/N.A (2-[18F]FDG) |
Pelvis |
0.3 ± 2.6 (Fat), − 0.0 ± 3.0 (soft tissue), − 0.9 ± 5.1 (bone) |
Dixon: 1.5 ± 6.5 Dixon: −0.3 ± 10.0 Dixon: −25.1 ± 12.7 |
| Arabi et al. [237] | Generation of CT-derived pseudo-μ-maps from PET TOF sinogram data using the 2D HighRes framework |
52/16/N.A (2-[18F]FDG) |
Brain | 2.9 ± 3.1 (head), 2.0 ± 10.6 (soft tissue), 1.2 ± 10.2 (bone) | N/A |
| Armanious et al. [259] | Generation of pseudo-CT from NAC PET 2-[18F]FDG images using a 2D GAN framework |
100/N.A/25 (2-[18F]FDG) |
Whole body | − 5.6 ± 7.2 (left lung), − 3 ± 11.7 (right lung) | N/A |
| Dong et al. [248] | Generation of AC from NAC PET images using a 2D cycleGAN framework |
25/1 (leave-one-out)/30 (2-[18F]FDG) |
Whole body | − 17.0 ± 12.0 (lung), 2.1 ± 2.5 (heart), 2.8 ± 5.2 (lesions) □ | N/A |
| Hu et al. [258] | Generation of pseudo-CT and AC from NAC PET images using a 2D Wasserstein GAN |
40/5/N.A (2-[18F]FDG) |
Whole Body | 6.4 ± 3.8 (brain), 4.4 ± 3.2 (heart), 4.3 ± 5.4 (lung) □ | N/A |
| Ladefoged et al. [215] | Similar to [212] but using DIXON images as input |
403 + 5 (transfer learning)/207/104 (2-[18F]FDG) |
Brain | ~ −0.3 | DixonSB: 0.8 ± 2.4 |
| Anaya et al. [303]*** | Generation of pseudo-CT from Dixon water images using the (2D) pix2pix framework |
9/2/1 (N/A) |
Head & neck | 2.1^ | N/A |
| Chen et al. [210] | Generation of pseudo-CTs from R1 images as derived from UTE using a residual 3D U-Net |
72/18/84 ([18F]Florbetapir) |
Brain | 0.1 ± 0.6 □ | N/A |
| Choi et al. [225] | Generation of tracer-specific pseudo-μ-map from MLAA images using a 3D U-Net |
60/20/20 (2-[18F]FDG) |
Brain | < 5 | N/A |
| Gong et al. [217] | Same as [214] but using images from a UTE/multi-echo sequence as input | 30/5/N.A ([11C]PiB, [18F]MK6240) | Brain | < 2 | N/A |
| Gong et al. [228] | Generation of the pseudo-CTs from Dixon images using a 3D cycleGAN |
28/4/N.A ([18F]FDG) |
Brain | ~ 3 | Dixon: ~ 8 |
| Hashimoto et al. [41]* | Generation of pseudo-CTs from NAC PET images using a 2D U-Net algorithm and mixed tracer training data |
1091/N.A/70 (6 tracers) |
Brain | − 5.7 ± 5.0 (2-[18F]FDG) | N/A |
| Kläser et al. [236] | Pseudo-CT generation from T1 & T2 images using HighRes3DNET and imitation learning |
16/4/23 (2-[18F]FDG) |
Brain | 4.04 ± 0.5 ^□ | N/A |
| Pozaruk et al. [243] | Generation of pseudo-CTs from Dixon images using a 2D cycleGAN framework |
18 /N.A/10 ([68 Ga]Ga-PSMA) |
Pelvis | 2.2 ^□ | Dixon: 10.3, DixonSB: 8.7 |
| Shiri et al. [305] | Generation of tracer and sight-specific AC from NAC PET images using a 2D U-Net and transfer learning techniques | 1110 (2-[18F]FDG) & 855 ([68 Ga]Ga-PSMA)/N.A./95 ([68 Ga]Ga-PSMA) | Whole body | 2.72 ± 7.5 | N/A |
| Ahangari et al. [242] | Generation of pseudo-CT from Dixon images using a 3D U-Net | 11 (transfer learning)/N.A./15 | Whole body |
2.1 ± 2.4 (Brain), − 4.9 ± 12.1 (lung), − 4.0 ± 6.5 (bone) |
DixonSB: 2.1 ± 3.2, DixonSB: −4.3 ± 20.3, DixonSB: −7 ± 12.4 |
| Arabi et al. [271] | Pseudo-CT generation from in-phase Dixon images using the 2D HighResNet |
20/5/N.A (2-[18F]FDG) |
Whole body | − 3.7 ± 5.5 (lung), 1.1 ± 3.1 (bone), 2.1 ± 3.7 (cerebellum) | N/A |
| Hwang et al. [306] | Generation of tracer-specific pseudo-μ-map from MLAA images using a 3D U-Net | 60/20/20 (2-[18F]FDG) | Whole body |
1.2 ± 5.7 (lung lesions), 0.2 ± 3.8 (bone lesions) |
N/A |
| Olin et al. [211] | Generation of pseudo-CTs from Dixon images using a 3D U-Net | 800 heads + 17 head & neck (transfer learning)/leave-one-out/10 (2-[18F]FDG) | Head & neck | − 0.6 ± 2.0 (lesions) | DixonSB: −3.5 ± 4.6 |
| Sari et al. [289] | Air pocket segmentation from Dixon images using a 3D U-Net followed by generation of pseudo-CTs from Dixon images using a second 3D U-Net |
30/5/N.A (2-[18F]FDG) |
Pelvis | 2.6 | DixonSB: 5.1 |
| Toyonaga et al. [255] | Generation of tracer-specific pseudo-μ-map from MLAA images using a 3D U-Net |
40/22/73 (2-[18F]FDG) 40/22/36 ([68 Ga]DOTATATE) 40/22/50 ([18F]Fluciclovine) |
Whole body |
Thorax: − 1.5 ± 2.3 (2-[18F]FDG), 2.2 ± 2.3 ([18F]Dotatate), − 2.9 ± 1.8 ([18F]Fluciclovine) |
N/A |
| Wang et al. [307] | Generation of AC from NAC PET images using a 2D U-Net with deformable transformer layers |
21/4/5 (13N-Ammonia) |
Thorax | 10.1 ± 2.9 (myocardium) | N/A |
| Shiri et al. [275] | Training of a “global” model for multi-centre trials, by feeding sight-specific trained models to it. AC from NAC PET images using a 2D U-Net are generated |
180/60/60 (2-[18F]FDG) |
Whole body | -0.1 ± 0.1 | N/A |
The mean relative error along with the standard deviation (where available) in radiotracer uptake for the whole region is reported unless otherwise specified. CT was used for reconstructing the reference images unless otherwise specified
†Number of patients used for training, validating and testing the model
*Dixon Segbone method (DixonSB) [31]
**Transmission data used for reconstructing the reference PET images
***An atlas method [71] used for reconstructing the reference PET images, ^ relative absolute error is reported, □ voxel-wise error is reported