Table 1.
Articles included in the review focusing on image-to-image translation and cross-modality synthesis.
| Author | Year | Application | Population (No. of Patients) |
Imaging Modality | ML Model | Results |
|---|---|---|---|---|---|---|
| Jin | 2019 | Image-to-Image translation and cross-modality synthesis | 202 patients | MRI from CT image | MR-GAN | MAE: 19.36 PSNR: 65.35 SSIM: 0.25 |
| Kazemifar | 2019 | Image-to-Image translation and cross-modality synthesis | 66 patients | CT from MRI | GAN | mean absolute difference below 0.5% (0.3 Gy) |
| Dai | 2020 | Image-to-Image translation and cross-modality synthesis | 274 subjects (54 patients with low-grade glioma, and 220 patients with high-grade glioma) | MRI | multimodal MR image synthesis method unified generative adversarial network. | NMAEs for the generated T1c, T2, Flair: 0.034 ± 0.005, 0.041 ± 0.006, and 0.041 ± 0.006. PSNRs: 32.353 ± 2.525 dB, 30.016 ± 2.577 dB, and 29.091 ± 2.795 dB. SSIMs: 0.974 ± 0.059, 0.969 ± 0.059, and 0.959 ± 0.05. |
| Hamghalam | 2020 | Image-to-Image translation and cross-modality synthesis | Various datasets | MRI-HTC | Cycle-GAN | Dice similarity scores: 0.8%, (whole tumor) 0.6% (tumor core) 0.5% (enhancing tumor). |
| Maspero | 2020 | Image-to-Image translation and cross-modality synthesis | 60 pediatric patients | SynCT from T1-weighted MRI | cGANs | mean absolute error of 61 ± 14 HU pass-rate of 99.5 ± 0.8% and 99.2 ± 1.1% |
| Sanders | 2020 | Image-to-Image translation and cross-modality synthesis | 109 brain tumor patients | relative cerebral blood volume (rCBV) maps from computed from DSC MRI, from DCE MRI of brain tumors |
cGANs | Pearson correlation analysis showed strong correlation (ρ = 0.87, p < 0.05 and ρ = 0.86, p < 0.05). |
| Wang | 2020 | Image-to-Image translation and cross-modality synthesis | 20 patients | MRI-PET | cycleGANs | PSNR > 24.3 SSIM > 0.817 MSE ≤ 0.036. |
| Lan | 2021 | Image-to-Image translation and cross-modality synthesis | 265 subjects | PET-MRI | 3D self- attention conditional GAN (SC- GAN) constructed |
NRMSE:0.076 ± 0.017 PSNR: 32.14 ± 1.10 SSIM: 0.962 ± 0.008 |
| Bourbonne | 2021 | Image-to-Image translation and cross-modality synthesis | 184 patients with brain metastases | CT-MRI | 2D-GAN(2D U-Net) | mean global gamma analysis passing rate: 99.7% |
| Cheng | 2021 | Image-to-Image translation and cross-modality synthesis | 17 adults | Two-dimensional fMRI images using two-dimensional EEG images; |
BMT-GAN | MSE: 128.6233 PSNR: 27.0376 SSIM: 0.8627 VIF: 0.3575 IFC: 2.4794 |
| La Rosa | 2021 | Image-to-Image translation and cross-modality synthesis | 12 healthy controls and 44 patients diagnosed with Multiple Sclerosis | MRI (MP2RAGE uniform images (UNI) from MPRAGE) |
GAN | PSNR: 31.39 ± 0.96 NRMSE: 0.13 ± 0.01 SSIM: 0.98 ± 0.01 |
| Lin | 2021 | Image-to-Image translation and cross-modality synthesis | AD 362 subjects; 647 images CN 308 subjects; 707 images pMCI 183 subjects; 326 images sMCI 233 subjects; 396 images |
MRI-PET | Reversible Generative Adversarial Network (RevGAN) | Synthetic PET: PSNR: 29.42 SSIM: 0.8176 PSNR: 24.97 SSIM: 0.6746 |
| Liu | 2021 | Image-to-Image translation and cross-modality synthesis | 12 brain cancer patients | SynCT images from T1-weighted postgadolinium MR | GAN model with a residual network (ResNet) | Average gamma passing rates at 1%/1 mm and 2%/2 mm were 99.0 ± 1.5% and 99.9 ± 0.2%, |
| Tang | 2021 | Image-to-Image translation and cross-modality synthesis | 37 brain cancer patients | SynCT from T1-weighted MRI | GAN | Average gamma passing rates at 3%/3 mm and 2%/2 mm criteria were 99.76% and 97.25% |
| Uzunova | 2021 | Image-to-Image translation and cross-modality synthesis | Various datasets | MRI (T1/Flair to T2, healthy to pathological) | GAN | T1 → T2 SSIM: 0.911 MAE: 0.017 MSE: 0.003 PSNR: 26.0 Flair → T2 SSIM: 0.905 MAE: 0.021 MSE: 0.004 PSNR: 24.6 |
| Yang | 2021 | Image-to-Image translation and cross-modality synthesis | 9 subjects | Multimodal MRI-CT registration into monomodal sCT-CT registration | CAE-GAN | MAE: 99.32 |