Table 5.
Summary of reviewed studies using deep learning for lung image synthesis
| Study |
Modality
(original ⇒ target) |
Disease | Number of subjects | Dimensionality | Model | Preprocessing |
Percentage data split
(training*/testing) |
Performance |
|---|---|---|---|---|---|---|---|---|
| Bi et al. (2017) 118 | CT ⇒ FDG PET | Lung cancer | 50 | 2D | Multichannel-GAN (U-Net) | Manual segmentation of tumour/lymph nodes, axial slices containing tumours only | 50/50 | MAE = 4.6 PSNR = 28.06 |
| Jang et al. (2019) 119 | CT ⇒ 99 mTc-MAA SPECT perfusion | Lung cancer | 54 | 2D | Conditional GAN | Resized images, segmentation and removal of bone, soft tissue and heart | 91/9 | MS-SSIM = 0.87 γ index 2%/2mm = 97.7±1.2% |
| Zhong et al. (2019) 61 | 4DCT ⇒ CT ventilation | Lung cancer, COPD | 82 | 2D | Deep CNN | Images cropped to ROI | 10-fold CV | MSE = 7.6% γ index 5%/5mm = 80.6±1.4% SSIM = 0.880±0.035 |
| Liu et al. (2020) 43 | 4DCT ⇒ 99 mTc-Technegas SPECT ventilation | Lung cancer, oesophageal cancer | 50 | 2D | U-Net | Pre-computed lung mask, normalisation [0,1], post-processing normalisation [90th percentile] | 10-fold CV | Spearman’s ρ = 0.73±0.17 DSC = 0.73±0.09 |
| Ren et al. (2019) 29 | CT ⇒ 99 mTc-MAA SPECT perfusion | Lung cancer | 30 | 3D | U-Net | Clipped [-1000,–300 HU] for segmentation, normalisation [0,1] | 83/17 | Correlation coefficient = 0.53 ± 0.14 |
| Preiswerk et al. (2018) 120 | Ultrasound ⇒ MRI | NR | 7 | 3D | LRCN | PCA = 10 components | 66/33 (conducted in time segments) | SSE = 39.0 ± 12 |
| Olberg et al. (2018) 44 | MRI ⇒ CT | NR | 41 | NR | GAN (U-Net) | Normalisation [NR], pre-computed body mask | 90/10 | 3D γ index passing rate 99.2% Lung V20% difference = 0.11% |
CNN, Convolutional neural network; COPD, Chronic obstructive pulmonary disease; FDG, Fluorine-18‐fluorodeoxyglucose; GAN, Generative adversarial network; HU, Hounsfield unit; LRCN, Long-term recurrent convolutional network; MAE, Mean absolute error; MSE, Mean square error; MS-SSIM, Multi-scale structural similarity index metric; NR, Not reported; PCA, Principle component analysis; PSNR, Peak signal to noise ratio; ROI, Region of interest; SSE, Sum of squared error; 99mTc-MAA, Technetium-99m macroaggregated albumin.