Table 1.

Literature reviews of lung segmentation system using Hounsfield unit (HU) threshold, deformable boundaries, shape models, region/edge-based models, or machine learning (ML) based methods.

Study	Method	# Subjects	System Evaluation
Amato et al. [16,17]	1. Grey scale thresholding 2. Rolling ball algorithm.	17 CT patients.	The area under the ROC curve (AUC) of the system was $93\%$.
Hu et al. [13]	1. Grey scale thresholding. 2. Dynamic programming. 3. Morphological operations.	eight normal CT patients.	The average intrasubject change was $2.75\%\pm 2.29\%$.
Itai et al. [22]	1. Grey scale thresholding. 2. Active contour model.	9 CT Patients.	Qualitative evaluation only.
Silveria et al. [23,24]	1. Grey scale thresholding. 2. Geometric active contour. 3. Level sets. 4. Expectation-maximization (EM) algorithm.	Stack of chest CT slices.	Qualitative evaluation only.
Gao et al. [19]	1. Grey scale thresholding. 2. Anisotropic diffusion. 3. 3D region growing. 4. Dynamic programming. 5. Rolling ball algorithm.	eight CT scans.	The average overlap coefficient of the system was $99.46\%$.
Pu et al. [18]	1. Grey scale thresholding. 2. Geometric border marching.	20 CT patients.	Average over-segmentation and under-segmentation ratio were $0.43\%$ and $1.63\%$, respectively.
Korfiatis et al. [57]	1. k-means clustering 2. Support vector machine (SVM)	22 CT patients.	The mean overlap coefficient of the system was higher than $89\%$.
Wang et al. [58]	1. Gray scale thresholding. 2. 3D gray-level co-occurrence matrix (GLCM) [59,60].	76 CT patients.	The mean overlap coefficient of the system was $96.7\%$.
Van Rikxoort et al. [15]	1. Region growing. 2. Grey scale thresholding. 3. Dynamic programming. 4. 3D hole filling. 5. Morphological closing.	100 CT Patients.	The accuracy of the system was $77\%$.
Wei et al. [20]	1. Histogram analysis and connected-component labeling. 2. Wavelet transform. 3. Otsu’s algorithm.	nine CT patients.	The accuracy range of the system was $76.7-94.8\%$.
Ye et al. [21]	1. 3D fuzzy adaptive thresholding. 2. Expectation–maximization (EM) algorithm. 3. Antigeometric diffusion. 4. Volumetric shape index map. 5. Gaussian filter. 6. Dot map. 7. Weighted support vector machine (SVM) classification.	108 CT patients.	The average detection rate of the system was $90.2\%$.
Sun et al. [27]	1. Active shape model matching method. 2. Rib cage detection method. 3. Surface finding approach.	60 CT patients.	The Dice similarity coefficient (DSC) and mean absolute surface distance of the system were $97.5\%\pm 0.6\%$ and $0.84\pm 0.23$, respectively.
Sofka et al. [29]	1. Shape model. 2. Boundary detection.	260 CT patients.	The errors in segmenting left and right lung were $1.98\pm 0.62$ and $1.92\pm 0.73$, respectively.
Hua et al. [30]	Graph-based search algorithm.	19 pathological lung CT patients.	The sensitivity, specificity, and Hausdorff distance of the system were $98.6\%\pm 1.1\%$, $99.5\%\pm 0.3\%$, and $13.3\pm 4.7$, respectively.
Nakagomi et al. [61]	Min-cut graph algorithm.	97 CT patients	The sensitivity and Jaccard index of the system were $91.2\%\pm 13.3\%$, and $97.7\%\pm 1.1\%$, respectively.
Mansoor et al. [52]	1. Fuzzy connectedness segmentation algorithm. 2. Texture-based random forest classification. 3. Region-based and neighboring anatomy guided correction segmentation.	more than 400 CT patients.	The DSC, Hausdorff distance, sensitivity, and specificity of the system were $95.95\%\pm 0.34\%$, $19.65\pm 12.84$, $96.84\%\pm 1.63\%$, and $92.97\%\pm 0.68\%$, respectively.
Yan et al. [62]	Convolution neural network (CNN).	861 CT COVID-19 patients.	The system achieved DSC of $98.7\%$ and $72.6\%$, sensitivity of $98.6\%$ and $75.1\%$, and specificity of $99\%$ and $72.6\%$ for normal and COVID-19-infected lung, respectively.
Fan et al. [63]	1. COVID-19-infected lung segmentation convolution neural network (Inf-Net). 2. Semi-supervised Inf-Net (Semi-Inf-Net).	100 CT images.	The DSC (sensitivity, specificity) of Inf-Net and Semi-Inf-Net were $68.2\%$ ($69.2\%$, $94.3\%$) and $73.9\%$ ($72.5\%$, $96\%$), respectively.
Oulefki et al. [64]	Multi-level entropy-based threshold approach.	297 CT COVID-19 patients.	The DSC, sensitivity, specificity, and precision of the system were $71.4\%$, $73.3\%$, $99.4\%$, and $73.9\%$, respectively.
Sharafeldeen et al. [65]	1. Linear combination of Gaussian. 2. Expectation-maximization (EM) algorithm. 3. Modified k-means clustering approach. 4. 3D MGRF-based morphological constraints.	32 CT COVID-19 patients.	The Overlap coefficient, DSC, absolute lung volume difference (ALVD), and 95th-percentile bidirectional Hausdorff distance (BHD) were $91.76\%\pm 3.29\%$, $95.67\%\pm 1.83\%$, $2.93\pm 2.39$, and $4.86\pm 5.01$, respectively.
Zhao et al. [66]	1. Grey scale thresholding. 2. 3D V-Net. 3. Deformation module.	112 CT patients.	DSC, sensitivity, specificity, and mean surface distance error of the system were $97.96\%$, $98.4\%$, $99.54\%$, and $0.0318$, respectively.
Sousa et al. [67]	Hybrid deep learning model, consisted of U-Net [68] and ResNet-34 [69] architectures.	385 CT patients, collected from five different datasets.	The mean DSC of the system was higher than $93\%$, and the average Hausdorff distance was less than $5.2$.
Kim et al. [70]	Otsu’s algorithm.	447 CT patients.	Sensitivity, specificity, accuracy, AUC, and F1-score of the system were $96.2\%$, $97.5\%$, $97\%$, $96.8\%$, and $96.1\%$, respectively.