Table 1 Summary of COPD identification and staging by different teams

2013

[64]

Mets et al.

1140 inspiratory and expiratory CT

COPD identification

- The segmentation of lung and airway

ACC=82.8%, SEN=88%, SPE=73.2%, PPV=80.2%, NPV=84.2%

437 COPD

- Three quantitative CT biomarkers (emphysema, air trapping, and bronchial wall thickness)

- Logistic regression

2016

[65]

Ying et al.

COPDGene

COPD GOLD classification (GOLD 2011)

- Fisher score method

ACC=97.2%

- Deep belief network

2017

[66]

González et al.

COPDGene: 3881 COPD vs 4387 non-COPD, ECLIPSE: 1727 COPD

COPD identification and stage

- Four canonical views of the CT scan

Identification:

- CNN with three convolutional layers

(COPDGene) ACC=77.3% AUC= 0.856

(ECLIPSE) AUC=0.548

Stage:

(COPDGene) ACC=51.1%

(ECLIPSE) ACC=29.4%

2017

[67]

Cheplygina et al.

-Danish Lung Cancer Screening (DLCST)

COPD identification

- Gaussian texture features

DLCST: AUC=0.684

-Partial COPDGene

- Multiple instance learning

COPDGene: AUC=0.962

- Transfer learning

Frederikshavn AUC=0.969

2018

[68]

Sathiya et al.

—

COPD identification

- Gray Level Co-occurrence Matrix

—

- Fuzzy c-means clustering

- CNN classifier

2020

[69]

Singla et al.

COPDGene

GOLD diagnosis and stage

- Discriminative network

Diagnosis: AUC=0.82 Recall=0.80

- Attention mechanism

Stage: ACC=65.44%

- Generative network

2020

[70]

Xu et al.

190 COPD vs 90 non-COPD

COPD identification

- CT lung image

ACC=99.29%, AUC=0.9826, SEN=99.47%

- The fourth convolutional layer for 9 feature extraction Principle component analysis MIL, Citation-KNN for classification

SPE=98.89% F1-score=0.9947

2020

[71]

Du et al.

190 COPD vs 90 non-COPD

COPD identification

- Snapshots of different view of airway tree

ACC=88.6%

- CNN with Bayesian optimization

- Majority voting

2020

[72]

Tang et al.

1304 COPD vs 1285 non-COPD (PanCAN)

COPD identification

- CT images

PanCAN: AUC=0.889

- ResNet 152

ECLIPSE: AUC=0.886

PPV=0.847, NPV=0.755

2021

[73]

Ho et al.

204 COPD vs 392 non-COPD

COPD identification

- Parametric-response mapping of CT image

ACC=89.3%, SEN=88.3%, AUC=0.937

- 3D CNN

2021

[74]

Hasenstab et al.

COPDGene.

GOLD stage

- Co-registration

SEN=88.25%, SPE=74.5%, AUC=0.905

- Lung segmentation

- Measurements of emphysema and air trapping

- Logistic regression

2022

[75]

Chen et al.

707 COPD vs 4116 non-COPD

COPD identification

- Graph convolutional MIL

AUC=0.96

- Adaptive additive margin loss

2022

[76]

Li et al

249 patients with stable COPD and 73 controls

COPD identification and stage

- Radiomics

Identification:

- Feature selection methods, including variance threshold, Select K Best method, and least absolute shrinkage and selection operator (LASSO)

ACC=0.941, SEN=0.940

- Support vector machine (SVM)

SPE=0.955, AUC=0.970

Stage:

ACC=0.759, SEN=0.723

SPE=0.805, AUC=0.799

2022

[77]

Makimoto et al.

602 COPD, 602 non-COPD

COPD identification

- Resampling, segmenting the lung and removing the airways

AUC=0.78

- Radiomics

- Elastic Net (feature selection)

- SVM

2022

[78]

Zhang et al.

204 COPD vs 392 non-COPD

COPD identification

- Lung parenchyma and bronchial wall patch of CT lung images

ACC=81.7%, SEN=81.0%

- DenseNet-201

SPE=81.6%, AUC=0.899

2022

[37]

Wu et al.

291 COPD vs 290 non-COPD

COPD identification

- Snapshots of different view of airway tree and lung field

ACC=94.7%

- ResNet-26

SEN=92.9%

- Majority voting

SPE=96.7%

2022

[38]

Sun et al.

749 non-COPD vs 644 COPD

COPD identification and stage

- CT images

Identification

- ResNet18, attention MIL

SEN=80.5%, SPE=92.5%, AUC=0.934

- Multi-channel 3D residual network

Stage:

SEN=76.5%, SPE=92.2%, AUC=0.912

2022

[79]

Li et al.

204 COPD vs 392 non-COPD

COPD identification

- Segmented lung parenchyma of CT images

ACC=77%, Precision=0.80

- Graph convolutional network

F1-score=0.78, AUC=0.81

2022

[80]

Yang et al.

468 subjects with Stage 0 to IV

COPD identification and stage

- Radiomics

ACC=0.80, Precision=0.943

- LASSO

F1-score=0.946, AUC=0.94

- Multi-layer perceptron

2022

[81]

Yang et al.

465 subjects with Stage 0 to IV (129, 108, 121, and 107)

COPD identification and stage

- Radiomics, 3D CNN features

ACC=0.943, Precision=0.943

- Auto-metric graph neural network

F1-score=0.946, AUC=0.984

2023

[32]

Wu et al.

271 COPD vs 290 non-COPD

COPD identification

- CT images, Snapshots of different view of airway tree and lung field

ACC=95.8%

- Attention MIL, LR

SEN=95.3%

SPE=96.5%

2023

[82]

Almeida et al.

COPDGene and COSYCONET [83]

COPD identification

- Spatial alignment, Lung. Traches, and aorta segmentation (Pre-processing)

COPDGene:

- Self-supervised learning

AUC=0.843

- Generative Model

COSYCONET:

AUC=0.679

2023

[84]

Xue et al.

363 COPD vs 437 non-COPD

COPD identification

- Transfer learning (pre-trained Resnet-50)

ACC=92%, SEN=92%

- Pseudo-color method

SPE=91.95%, AUC=0.9544

- Two stage attention MIL

2023

[85]

Zhou et al.

COPD (4,912),

COPD identification

- Multi-modal (Radiograph, chief complaint, and demographics and lab test results)

AUC=0.922

- Transformer-based representation and classification

2023

[86]

Puchakayala et al.

COPDGene

COPD identification

- Demographics features, emphysema and radiomics features of CT images

Standard-Dose CT Data:

- CatBoost

PPV=0.86, NPV=0.83, AUC=0.90

Low-Dose CT Data:

PPV=0.79, NPV=0.80, AUC=0.88

2023

[87]

Yu et al.

COPDGene

COPD stage

- Self-supervised Learning

ACC=0.65

- CNN