Table 2.
Summary of reviewed studies on deep learning for lung image segmentation. The entries are arranged alphabetically by pulmonary region of interest (ROI), followed by modality
| Study | Modality | ROI | Disease | Number of subjects | Dimentionality | Architecture | Pre-processing |
Percentage data split
(training*/testing) |
Performance |
|---|---|---|---|---|---|---|---|---|---|
| Wang et al. (2018) 13 | CT | Whole lung | COPD, IPF | 575 | 2D | ResNet-101 | Clipped −1000 to +1000 HU, Normalisation [0,1] | 5-fold CV | DSC = 0.988 ± 0.012 ASD = 0.562±0.52 mm |
| Dong et al. (2019) 70 | CT | Whole lung | Lung cancer | 35 | 3D | U-Net-GAN | LOOCV | DSC = 0.97±0.01 HD95 = 2.29±2.64 mm MSD = 0.63±0.63 mm |
|
| Liu et al. (2019) 30 | CT | Whole lung | NR | 100 | 2D | SegNet | Class grouping, Normalisation [−1000,800] | 40/60 | DSC = 0.98 |
| Lustberg et al. (2018) 71 | CT | Whole lung | Lung cancer | 470 | NR | CNN | 95/5 | DSC = 0.99±0.01 Median HD = 0.4±0.2 cm |
|
| Negahdar et al. (2018) 12 | CT | Whole lung | Multiple | 83 | 3D | V-Net | Bounding box for lung, cropped to bounding box | 58/42 | DSC(n = 12)=0.983±0.002 DSC(n = 23)=0.990±0.002 |
| Soans & Shackleford (2018) 48 | CT | Whole lung | Lung cancer | 422 | 3D | CNN with spatial constraints | ROI extraction for organ localisation | 71/29 | ROC(Left)=0.954 ROC(right)=0.949 |
| Soliman et al. (2018) 72 | CT | Whole lung | NR | 95 | 3D | Deep-CNN | Post-processed hole filling | LOOCV | DSC = 0.984±0.068 HD95 = 2.79±1.32 mm PVD = 3.94±2.11% |
| Sousa et al. (2019) 14 | CT | Whole lung | Lung lesion | 908 | 3D | Modified V-Net | Clipped [−1000, 400 HU] | 98/2 | ASD = 0.576 mm DSC = 0.987 |
| X. Zhou et al. (2017) 73 | CT | Whole lung | NR | 106 | 2D/3D | FCN VGG16 | Transfer learning from ImageNet ILSVRC‐2014 | 95/5 | JSC = 0.903±0.037 |
| Zhu et al. (2019) 49 | CT | Whole lung | Lung Cancer | 66 | 3D | U-Net | Cropping to ROI | 55/45 | DSC = 0.95±0.01 MSD = 1.93±0.51 mm HD95 = 7.96±2.57 mm |
| Gerard et al. (2018) 74 | CT | Whole lung | COPD, IPF | 1749 | 3D | Course-Fine ConvNet | Transfer learning from COPDGene and SPIROMICS, fine-tuned on animal model | 92/8 | JSC = 0.99 ASD = 0.29 mm |
| Javaid et al. (2018) 15 | CT | Whole lung | Lung cancer | 13 | 2D | Dilated U-Net | Only axial slices selected, clipped −1000 to 3000 HU, Normalisation [0,1] | 94/6 | DSC = 0.99 ± 0.01 HD ≈ 4.5 mm |
| J. Xu & Liu (2017) 45 | CT | Whole lung | NR | 20 | 2D | MFCNN | gaussian denoising | 50/50 | DSC = 0.754 |
| Hu et al. (2020) 75 | CT | Whole lung | NR | 75 | 2D | Mask R-CNN +k-means | NR | DSC = 0.973 ±0.032 | |
| Hofmanninger et al. (2020) 16 | CT | Whole lung | Multiple | 266 | 2D | U-Net | Body mask, Clipped [−1024, 600 HU], Normalisation [0,1] | 87/13 | DSC = 0.98 ±0.03 HD95 = 3.14 ±7.4 mm MSD = 0.62 ±0.93 |
| Xu et al. (2019) 76 | CT | Whole lung | Lung cancer, COPD | 224 | 2D | one layer CNN | Post-processed hole filling | 8-fold CV | DSC = 0.967 ±0.001 HD = 1.44±0.04 mm |
| Tustison et al. (2019) 47 | HP gas MRI Proton MRI |
Functional lung Whole lung |
NR NR |
113 268 |
2D 3D |
U-Net U-Net |
Template-based data augmentation, N4 bias correction, denoising | 65/35 77/23 |
DSC (HP gas)=0.92 DSC (Proton) = 0.94 |
| Akila Agnes et al. (2018) 31 | LDCT | Whole lung | NR | 220 | 2D | CDWN | Normalised [mean = 0] | 91/9 | DSC = 0.95 ± 0.03 JSC = 0.91 ± 0.04 |
| Zha et al. (2019) 46 | UTE proton MRI | Whole lung | Healthy, CF, asthma | 45 | 2D | CED (U-Net and autoencoder) | Denoising, bias field correction, body mask | 5-fold CV | DSC (right) = 0.97±0.015 DSC (left) = 0.96±0.012 |
| Hwang & Park (2017) 77 | X-ray | Whole lung | Healthy, lung nodules | 247 | 2D | U-Net | 2-fold CV | DSC = 0.980±0.008 JSC = 0.961±0.015 ASD (mm) = 0.675±0.122 ACD (mm) = 1.237±0.702 |
|
| Souza et al. (2019) 78 | X-ray | Whole lung | Healthy, Tuberculosis | 138 | 2D | ResNet-18 with FC layer | Scaled to same input size, post processing erosion, dilation, filtering | 73/27 | DSC = 0.936 JSC = 0.881 |
| Dai et al. (2018) 64 | X-ray | Whole lung | Healthy, Tuberculosis, lung nodules | 385 | 2D | SCAN (structure correcting adversieral network) | Scaled to same input size | 85/15 | IoU = 94.7±0.4% DSC = 0. 973 ± 0.02 |
| C. Wang (2017) 79 | X-ray | Whole lung | Healthy, lung nodules | 247 | 2D | Multi task U-Net | Scaled to same input size, post processing hole filling | NR | JSC = 0.959 ± 0.017 AD = 1.29 ± 0.80 mm |
| Novikov et al. (2018) 32 | X-ray | Whole lung | Healthy, lung nodules | 247 | 2D | InvertedNet + All-dropout | Normalised [mean = 0, SD = 0] | 3-fold CV | DSC = 0.974 JSC = 0.949 |
| Hooda et al. (2018) 50 | X-ray | Whole lung | Healthy, Tuberculosis, lung nodules | 385 | 2D | FCN-8+dropout | Scaled to same input size, random cropping | 75/25 | DSC = 0.959 |
| Mittal et al. (2018) 51 | X-ray | Whole lung | Healthy, Tuberculosis, lung nodules | 385 | 2D | LF-SegNet | Scaled to same input size, random cropping | 48/52 | DSC = 0.951 |
| Gaal et al. (2020) 33 | X-ray | Whole lung | Healthy, Tuberculosis, lung nodules | 1047 | 2D | Adversarial attention U-Net | Scaled to same input size, CLAHE, Normalisation [−1,1] | 24/76 | DSC = 0.962±0.04 |
| Chen et al. (2019) 80 | CT | Lung tumour | Lung cancer | 134 | 3D | HSN (2D + 3D CNN) | 78/22 | DSC = 0.888±0.033 | |
| Jiang et al. (2018) 11 | CT, MRI | Lung tumour | Lung cancer | 400 CT (377) MRI (23) |
2D | Tumour aware semi-supervised Cycle-GAN | Scaled to same input size, Image synthesis from CT to MRI, body mask | 98/2 | DSC = 0.63 ± 0.24 HD95 = 11.65±6.53 |
| Jiang et al. (2019) 17 | CT, MRI | Lung tumour | Lung cancer | 405 CT (377) MRI (28) |
2D | Tumour aware pseudo MR and T2w MR U-Net | Scaled to same input size, Image synthesis from CT to MR, Clipped [−1000,500 HU] and [0,667], Normalised [−1, 1] | 95/5 | DSC = 0.75±0.12 HD95 = 9.36±6.00 mm VR = 0.19±0.15 |
| Tahmasebi et al. (2018) 18 | MRI | Lung tumour | Lung cancer | 6 | 2D | Adapted FCN | Rescaled 10–95% of intensities, Normalisation [0,1] | 5-fold CV | DSC = 0.91 ± 0.03 HD = 2.88 ± 0.86 mm RMSE = 1.20 ± 0.34 |
| Z. Zhong et al. (2019) 19 | FDG PET, CT | Lung tumour | Lung cancer | 60 PET (60) CT (60) |
3D | DFCN Co-Seg U-Net | Scaled to same input size, Clipped [−500,200 HU] and [0.01,20] | 80/20 | DSC (CT) = 0.861±0.037 DSC (PET) = 0.828±0.087 |
| Zhao et al. (2019) 52 | PET, CT | Lung tumour | Lung cancer | 84 PET (84) CT (84) |
3D | V-Net +feature fusion | Cropped to ROI | 57/43 | DSC = 0.85±0.08 VE = 0.15±0.14 |
| Zhou et al. (2019) 20 | CT | Lung tumour | NR | 1350 | 3D | P-SiBA | Transfer learning from ImageNet ILSVRC‐2014, Cropped to ROI, Rescaled by +1000 HU and dividing by 3000 and Normalisation [0,1] | NR | DSC = 0.809 ± 0.12 HD = 7.612 ± 5.03 mm vs = 0.883 ± 0.13 |
| Moriya et al. (2018) 53 | Micro CT | Lung tumour | Lung cancer | 3 | 3D | JULE CNN + k-means | Body mask, patch extraction | NMI = 0.390 | |
| Imran et al. (2019) 81 | CT | Lobes | COPD, ILD | 563 | 3D | Progressive dense V-Net | 48/52 | DSC (n = 84)=0.939±0.02 DSC (n = 154)=0.950±0.007 DSC (n = 55)=0.934 |
|
| Park et al. (2019) 21 | CT | Lobes | COPD | 196 | 3D | U-Net | Clipped [-1024,–400 HU] | 80/20 | DSC = 0.956 ± 0.022 JSC = 0.917 ± 0.031 MSD = 1.315 ± 0.563 HSD = 27.89±7.50 |
| Wang et al. (2018) 13 | CT | Lobes | COPD, IPF | 1280 | 3D | DenseNet | Clipped −1000 to +1000 HU, Normalisation [0,1] | 5-fold CV | DSC = 0.959±0.087 ASD = 0.873±0.61 mm |
| Hatamizadeh et al. (2019) 34 | CT | Lung lesion | NR | 87 | 3D | DALS CNN | Scaled to same input size, Normalisation [NR] | 90/10 | DSC = 0.869 ± 0.113 HD = 2.095 ± 0.623 mm |
| Kalinovsky et al. (2017) 54 | CT | Lung lesion | Tuberculosis | 338 | 2D | GoogLeNet CNN | Images cropped into four quadrants | 80/20 | IoU = 0.95 ROC = 0.775 |
| Gerard et al. (2019) 22 | CT | Lung fissure | COPD, Lung cancer | 5327 | 3D | Two Seg3DNets | Clipped [-1024,–200 HU], Linear rescaling | 30/70 | ASD = 1.25 SDSD = 2.87 |
| Sandkühler et al. (2019) 35 | MRI | Lung defect region | NR | 35 | 2D | GAE-LAE RNN with LCI Loss | Z-normalisation [−4,4], Lung mask, Normalisation [0,1], Histogram stretching | 80/20 | Qualitative evaluation - 42% images rated ‘very good’, 19% rated ‘perfect’ |
| Vakalopoulou et al. (2018) 82 | CT | ILD pattern | ILD | 46 | 2D | AtlasNet | 37/63 | DSC = 0.677 HD = 3.981 mm ASD = 1.274 mm |
|
| Anthimopoulos et al. (2019) 55 | CT | ILD pattern | ILD | 172 | 2D | FCN-CNN | Pre-computed lung mask | 5-fold CV | Accuracy = 81.8% |
| B. Park et al. (2019) 83 | CT | ILD pattern | COP, UIP, NSIP | 647 | 2D | U-Net | 88/12 | DSC = 0. 988 ± 0.006 JSC = 0.978 ± 0.011 MSD = 0.27 ± 0.18 mm HSD = 25.47 ± 13.63 mm |
|
| Gao et al. (2016) 56 | CT | ILD pattern | ILD | 17 | 2D | CNN based CRF unary classifier | Transfer learning from ImageNet, Pre-computed lung mask | Accuracy = 92.8% | |
| Suzuki et al. (2020) 84 | CT | Diffuse lung disease | NR | 372 | 3D | U-Net | 5-fold CV | DSC = 0.780±0.169 | |
| Wang et al. (2018) 85 | MRI | Foetal lung | NR | 18 | 2D | BIFSeg P-Net | Trained on different organs, Image specific fine-tuning | 66/33 | DSC = 0.854±0.059 |
| Rajchl et al. (2017) 36 | MRI | Foetal lung | Healthy, IUGR | 55 | 3D | DeepCut CNN + CRF | Bounding box for ROI, Bias correction, Normalisation [mean = 0], Transfer learning from LeNet | 5-fold CV | DSC = 0.749±0.067 |
| Edmunds et al. (2019) 86 | Cone-beam CT | Diaphragm | Lung cancer | 10 | 2D | Mask R-CNN | Scaled to same input size | 9-fold CV | Mean error = 4.4 mm |
| C. Wang et al. (2019) 57 | CT | Airways | NR | 38 | 3D | Spatial-CNN (U-Net) | Random cropping | 92/8 3-fold MCCV | DSC = 0. 887 ± 0.012 CO = 0.766 ± 0.06 |
| Juarez et al. (2019) 58 | CT | Airways | Lung cancer | 32 | 3D | U-Net GNN | Bounding box for ROI | 63/37 | DSC = 0.885 Airway completeness = 74% |
| Yun et al. (2019) 23 | CT | Airways | COPD | 89 | 2D | 2.5D CNN | Clipped [−700,700 HU] | 78/22 | Mean Branch detected = 65.7% |
| Juarez et al. (2018) 59 | CT | Airways | Healthy, CF, CVID | 24 | 3D | U-Net | Bounding box for ROI | 75/25 | DSC = 0.8 |
ACD, Average contour distance; AD, Average distance; ASD, Average surface distance; CDWN, Convolutional deep wide network; CE, Classification error; CF, Cystic fibrosis; CLAHE, Contrast limited adaptive histogram equalisation; CNN, Convolutional neural network; CO, Centreline overlap; COPD, Chronic obstructive pulmonary disorder; CV, Cross-validation; CVID, Common variable immunodeficiency disorders; DSC, Dice similarity coefficient; FDG, Fluorine-18‐fluorodeoxyglucose; GAN, Generative adversarial network; HD95, Hausdorff distance 95%; HD, Hausdorff distance; HSD, Hausdorff surface distance; HU, Hounsfield unit; ILD, Interstitial lung disease; IPF, Idiopathic pulmonary fibrosis; IUGR, Intrauterine growth restriction; IoU, Intersection over union; JSC, Jaccard similarity coefficient; LOOCV, Leave-one-out cross-validation; MAP, Mean average precision; MCCV, Monte carlo cross-validation; MSD, Mean surface distance; NMI, Normalised mutual information; NR, Not reported; NSIP, Nonspecific interstitial pneumonia; PVD, Percent ventilated defect; RMSE, Root mean square error; ROC, Receiver operating characteristic; ROI, Region of interest; SD, Standard deviation; SDSD, Standard deviation of surface distances; UIP, Usual interstitial pneumonia; VE, Volume error; VR, Relative volume ratio; VS, Volumetric similarity.
The entries are arranged alphabetically by pulmonary ROI, followed by modality.
a
The training data set includes internal validation data.