Table 1 Summary of related work on lung segmentation techniques for chest radiographs

[33]

- 230 chest radiographs

Overlap score:

- ASM with optimal local features

- ASM right: 0.882 ± 0.074

- Computationally complex

- find optimal displacements for landmarks using a non-linear kNN classifier instead of linear Mahalanobis distance

- ASM left: 0.861 ± 0.109

- Suffers the drawback of ASM

- ASM-OF right: 0.929 ± 0.026

- ASM-OF left: 0.887 ± 0.114

[17]

- JSRT dataset (247 images)

Various methods were compared

Overlap score:

- Highly supervised and required training

- Hybrid voting

- Hybrid voting: 0.949 ± 0.020

- PC postprocessed

- PC postprocessed: 0.945 ± 0.022

- Hybrid ASM-PC

- Hybrid ASM-PC: 0.934 ± 0.037

- Hybrid AAM-PC

- Hybrid AAM-PC: 0.933 ± 0.026

- ASM-tuned

- ASM-tuned: 0.903 ± 0.057

- AAM

- AAM: 0.847 ± 0.095

- Mean Shape

- Mean Shape: 0.713 ± 0.075

[10]

- 24 chest radiographs from portable device, all with pulmonary bacterial infections manifested as consolidations

- based on Bezier interpolation of salient control points

Sensitivity: 95.3%

- Lack of images

Specificity: 94.3%

[11]

- 58 chest radiographs from portable device, all with pulmonary bacterial infections manifested as consolidations

- Gray-level selective thresholding followed by ASM

Accuracy presented in a graph, between 92.5% - 94%.

- Lack of images

- Suffers the drawback of ASM

[15]

- 52 selected images from JSRT dataset

- Gaussian kernel-based fuzzy clustering algorithm with spatial constraints

Accuracy:

- Lack of images (only 52 were selected out of 247 images in JSRT dataset)

- 0.978 ± 0.0213

[13]

Dice similarity:

- Requires training and optimization

- 1,130 images

- rule-based method (thresholding, morphology and connected components) used to generate a seed mask

- 0.88 ± 0.07

- 400 from Shanghai Pulmonary Hospital (200 normal, 200 with pneumoconiosis)

- Using optimized canny edge parameters to detect the corner (costophrenic angle)

- 730 from different clinical sites in China (with normal and various pulmonary conditions)

[36]

- JSRT dataset (247 images)

Overlap score:

- Requires optimization and testing

- Fusing shape information with statistical model of the lungs’ shape

- 22 landmarks: 0.92 ± 0.063

- intensity-based iterative thresholding

- 28 landmarks: 0.94 ± 0.053

- optimization using ASM

[34]

- JSRT dataset (247 images)

- ASM for the lung segmentation, with bone detection algorithm

- Sensitivity: 0.956

- Suffers the drawback of ASM

- Specificity: 0.984

[14]

- JSRT dataset (247 images)

Accuracy:

- based on spatial relationships between lung structures, represented as fuzzy subsets of the image space

- Left axillary: 82.1%

- Need to label the lung structures

- segment the lung structures

- Right axillary: 85.2%

- Accuracy or overlap score of whole lung is not provided

- Left parahilar: 84.4%

- Right parahilar: 82.8%

- Left Paracardiac: 68.8%

- Right Paracardiac: 86.5%

- Left Basal: 81.5%

- Right Basal: 81.7%

[35]

Accuracy:

- Requires shape-learning stage

- JSRT dataset (93 normal images)

- Global edge and region force (ERF) field based ASM (ERF-ASM)

- JSRT left: 0.952 ± 0.013

- CXR from University of Alberta Hospital dataset (50 images with tuberculosis)

- PCA analysis to learn the lung fields’ shape

- JSRT right: 0.955 ± 0.014

- CXR left: 0.946 ± 0.015

- CXR right: 0.953 ± 0.017

[37]

3 stages:

Overlap score:

- JSRT dataset (247 images)

1. CBIR approach to identify small set of lung CXR using Radon transform with Bhattacharyya similarity measure

- JSRT: 0.954

- Need to be highly trained

- Montgomery dataset (138 images – 80 normal, 58 with tuberculosis)

2. Construction of patient-specific lung atlas

- Montgomery: 0.941

- Computationally complex

- India dataset (200 images – 100 normal, 100 abnormal)

3. Lung segmentation using graph cuts discrete optimization approach

- India: 0.917