Table 4.
Summary of reviewed studies using deep learning for lung image reconstruction
| Study | Modality | Disease | Number of patients | Dimensionality | Architecture | Preprocessing |
Percentage data split
(training*/testing) |
Performance |
|---|---|---|---|---|---|---|---|---|
| Beaudry et al. (2019) 41 | 4D cone beam CT | Lung cancer | 16 | 2D | Sino-Net (Modified U-Net) | Cropped to same input size, Sinogram Normalisation [0,1] | 88/12 | RMSE Translational = 1.67 mm (other metrics given) |
| Lee et al. (2019) 107 | CT | COPD | 60 | 2D | FCN | No sinogram used | Dataset 1: 80/20 Dataset 2: 40/60 |
Mean reduction RMSE (Dataset 1) = 65.7±15.8% Mean reduction RMSE (Dataset 2) = 59.6±5.5% |
| Ge et al. (2020) 108 | CT | Liver lesion | 5413 | 2D | ADAPTIVE-NET CNN | Convert from HU to linear attenuation coefficient | 90/10 | PSNR = 43.15±1.9 SSIM = 0.968±0.013 Normalized RMSE = 0.0071±0.002 |
| Duan et al. (2019) 42 | HP Gas MRI | COPD, nodule, PTB, healthy, asthma | 72 | 2D | C-Net and F-Net (U-Net based) | Under sampled K-space (AF = 4), Removed SNR below 6.6, Normalisation [0,1] | NR | MAE = 4.35% SSIM = 0.7558 VDP bias = 0.01±0.91% |
| Dietze et al. (2019) 109 | 99 mTc-MAA SPECT | Liver Cancer | 128 | 2D | CNN | Initial filtered back projection | 94/6 | LSF = 5.1% CNR = 12.5 |
CNN, Convolutional neural network; CNR, Contrast to noise ratio; COPD, Chronic obstructive pulmonary disorder; EIT, Electrical impedance tomography; HU, Hounsfield unit; LSF, Lung shunting fraction; MAE, Mean absolute error; PSNR, Peak signal to noise ratio; PTB, Pulmonary tuberculosis; RMSE, Root mean square error; SSIM, Structural similarity index metric; VDP, Ventilation defect percentage; VDP, Volume defect percentage; 99mTc-MAA, Technetium-99m macroaggregated albumin.
a
The training data set includes internal validation data