Table 2.
Summary of artificial intelligence-based prediction model characteristics described in included studies.
| Study ID |
ROI | Segmentation Style |
AI Method | Labeling Procedure |
Pre-Processing | Augmentations | Model Structure |
Loss Function |
Comparison between algorithms | AI vs. Radiologist |
|---|---|---|---|---|---|---|---|---|---|---|
| Ardakani 2020 | Regions of infections | 2D | DL | by a radiologist with more than 15 years of experience in thoracic imaging | Manual ROI extraction by cropping, Normalization, transfer-learning |
NA | Ten well-known CNN |
NA | Ten well-known CNN |
Yes |
| Ardakani 2021 | CT chest | 2D | ML | By two radiologists | feature extraction | random scaling shearing horizontal flip |
ensemble method | NA | DT, KNN, Naïve Bayes, SVM | Yes |
| Ali 2021 | Whole image | 2D | DL | NA | Normalization, transfer-learning | Horizontal, vertical flip, Zoom, Shift |
ResNet50, ResNet101, Res Net 152 | NA | ResNet50, ResNet101, Res Net 152 | No |
| Han2021 | CT slices | 2D | DL | using the labeled COVID-19 dataset | both labeled and unlabeled data can be used | random scaling random translation, random shearing, horizontal flip |
a semi-supervised deep neural network | standard cross entropy loss | Supervised learning | No |
| Di2020 | Infected lesions | 2D | ML | NA | extracted both regional and radiomics features, Segmentation | NA | UVHL | cross- entropy |
SVM, MLP, iHL, tHL | No |
| Bai 2020 | Lung regions |
2D | DL | Lesions (COVID-19 or pneumo- nia) were manually labeled by2 radiologists |
Normalized, Segmentation | flips, scaling, rotations, random brightness and contrast manipulations, random noise, and blurring |
DNN | NA | No | Yes |
| Panwar 2020 | Whole image | 2D | DL | NA | Filter, dimension reduction, deep transfer learning | Shear, Rotation Zoom, shift |
A DL and Grad-CAM | binary cross-entropy loss | No | No |
| Kang 2020 | Lesion region | 3D | ML | NA | Segmentation, Feature Extraction, Normalization |
NA | Structured Latent Multi-View Representation Learning |
Ross-entropy loss | LR，SVM,GNB, KNN, NN | No |
| Liu 2021 | Each pneumonia lesion | 3D | ML | By three experienced radiologists | Feature Extraction, Filters |
NA | LASSO regression | NA | No | Yes |
| Chen 2021 | Consolidation and ground- glass opacity lesions |
3D | ML | By fifteen radiologists | Feature Extraction, wavelet filters, Laplacian of Gaussian filters, Feature selection |
NA | SVM | NA | No | No |
| Song 2020 | CT images | 2D | DL | NA | semantic feature extraction | NA | BigBiGAN | NA | SVM, KNN | Yes |
| Sun 2020 | Infected lung regions |
3D | DL | NA | Feature extraction |
NA | AFS-DF | NA | LR, SVM, RF, NN | No |
| Wang 2021 | Pneumonia lesions | 3D/2D | ML | By four radiologists | manual segmentation, Feature extraction |
NA | Linear, LASSO, RF, KNN | NA | Linear, LASSO, RF, KNN | Yes |
| Zhou 2021 | Lesion regions | 2D | DL | annotated by 2 radiologists | Segmentation | randomly flipped, cropped | Trinary scheme(DL) | Binary cross- entropy loss |
Plain scheme(DL) | Yes |
| Azouji2021 | X-ray images | 2D | DL | NA | Resizing x-ray images, Contrast limited adaptive histogram equalization, Deep feature extraction, Deep feature fusion | Rotation, translation | LMPL classifier | hinge loss function | NaiveBayes, KNN, SVM,DT, AdaBoostM2, TotalBoost,RF, SoftMax,VGG-Net | No |
| Cardobi 2021 | Lung area | 3D | ML | NA | Segmentation, features extraction | NA | LASSO model | NA | No | No |
| Yang 2021 | Pneumonia lesion | 3D | ML | artificially delineated | Segmentation, features extraction | spatially resampled | SVM | NA | Sigmoid-SVM, Poly-SVM, Linear-SVM, RBF-SVM | No |
| Chikontwe 2021 | CT slices | 3D | DL | NA | Segmentation | random transformations, flipping |
DA-CMIL | NA | DeCoVNet, MIL, DeepAttentionMIL, JointMIL | No |
| Zhu 2021 | CT images | 3D | DL | NA | Segmentation, features extraction |
NA | GACDN | Binary cross entropy | SVM，KNN,NN | No |
| Xie 2020 | CT slices | 3D | DL | NA | Segmentation, extract 2D local features and 3D global features |
random horizontal flip, random rotation, random scale, random translation, and random elastic transformation | DNN | NA | No | Yes |
| Qi 2021 | Lung field | 3D | DL | NA | segmentation of the lung field, Extraction of deep features, Feature representation | Image rotation, reflection, and translation | DR-MIL | NA | MResNet-50-MIL, MmedicalNet, MResNet-50-MIL-max-pooling, MResNet-50-MIL-Noisy-AND-pooling, MResNet-50-Voting, MResNet-50-Montages | Yes |
| Wang 2020 | Lung area | 3D | DL | NA | fully automatic DL model to segment, normalization, convolutional filter | NA | DL | NA | No | No |
| Yang 2020 | Infection regions |
3D | DL | NA | Class Re-Sampling Strategies, Attention Mechanism | scaling | Dual-Sampling Attention Network | binary cross entropyloss | RN34 + US, Attention RN34 + US Attention RN34 + SS Attention RN34 + DS |
No |
| Wu 2020 | CT slices | 3D | DL | NA | segmentation | NA | Multi-view deep learning fusion model |
NA | Single-view model | No |
| Zhang 2021 | Major lesions | 3D | DL | NA | Segmentation Feature extraction, Feature selection, |
scaling | DL-MLP | NA | DL-SVM,DL-LR, DL-XGBoost | Yes |
| Xin 2021 | Lungs, lobes, and detected opacities | 2D | DL | Confirmed by 3 experienced radiologists and human auditing | Segmentation Feature extraction |
NA | LR, MLP, SVM, XGboost |
NA | LR, MLP, SVM, XGboost |
No |
| Guo 2020 | NR | NA | ML | by two radiologists | Segmentation Feature extraction |
NA | RF | NA | No | No |
| Fang2020 | Primary lesion | 3D/2D | ML | by two chest radiologists | Segmentation feature extraction, feature reduction and selection |
NA | LASSO regression | NA | No | No |
| Xia 2021 | Lung areas | 2D | DL | NA | Segmentation feature extraction |
random rotation, scale, transmit |
DNN | Categorical cross- entropy |
No | Yes (pulmonary physicians) |
| Huang2020 | Pneumonia lesion | 3D | ML | by two chest radiologists | Segmentation feature extraction, filter |
NA | Logistic model | NA | No | No |
| Wu 2021 | Maximal regions Involving inflammatory lesions | 2D | ML | by two radiologists | feature extraction, manually delineating | NA | RF | NA | No | No |
| Chen2021 | Lesion region |
2D | ML | by two radiologists | Segmentation feature extraction,feature dimensionality reduction |
NA | WSVM | NA | RF, SVM LASSO |
Yes |
Abbreviations:AFS-DF:adaptive feature selection guided deep forest;AI:artificial intelligence;BigBiGAN: bi-directional generative adversarial network; CT: Computed tomography; CXR: Chest X-Ray; CNN: Convolutional neural network;DA-CMIL: Dual Attention Contrastive multiple instance learning; DT: Decision tree; DNN: Deep Neural Networks; DR-MIL: deep represented multiple instance learning; DL: deep learning; RF: Random Forests; GNB: Gaussian-Naive-Bayes; Grad-CAM: Gradient Weighted Class Activation Mapping; GACDN: generative adversarial feature completion and diagnosis network; IHL:Inductive Hypergraph Learning; KNN: K-nearest neighbor; LR: Logistic-Regression; LASSO: least absolute shrinkage and selection operator; LMPL: large margin piecewise linear; ML: machine learning; MLA: Machine learning algorithms; MLP: Multilayer Perceptron; MERS: Middle East respiratory syndrome; NN: Neural-Networks; ROI: Region of interest; SVM: Support vector machine; THL: Transductive Hypergraph Learning; 2D: two-dimensional;3D: three-dimensional;UVHL: Uncertainty Vertex-weighted Hypergraph Learning; WSVM: weighted support vector machine