Table 3.
Review of the literature for TB detection using CXRs based on different feature extraction methods.
| References | Highlights | Pretraining | Dataset |
|---|---|---|---|
| Govindarajan and Swaminathan (2021) | Texture-based feature descriptors with ML classifier | No | Montgomery |
| Alfadhli et al. (2017) | Used SURF as feature extractor and SVM as classifier | No | Montgomery |
| Jaeger et al. (2014) | Used texture-based features (LBP, HOG) and statistical feature with ML Classifier | No | Shenzhen, Montgomery |
| Chandra et al. (2020) | Used shape and textural features with SVM | No | Shenzhen, Montgomery |
| Santosh et al. (2016) | Used PHOG as features with MLP as classifier | No | Shenzhen, Montgomery |
| Duong et al. (2021) | Used Pretrained EfficientNet and ViT, and developed a hybrid of two | Yes | Shenzhen, Montgomery, Chestxray14, COVID-CXR (Chowdhury et al., 2020) |
| Ayaz et al. (2021) | Used Feature ensemble of handcrafted and deep features | Yes | Shenzen, Montgomery |
| Dasanayaka and Dissanayake (2021) | Generated synthetic images, performed segmentation and used feature ensemble for classification | Yes | Shenzhen, Montgomery |
| Msonda et al. (2020) | Used spatial pyramid pooling for deep feature extraction | Yes | Shenzhen, Montgomery, private |
| Sahlol et al. (2020) | Used Meta-heuristic approach for Deep feature selection | Yes | Shenzen, Montgomery, PedPneumonia |
| Rahman et al. (2020b) | Performed segmentation and used different visualization techniques | Yes | Shenzhen, Montgomery, NIAID TB, RSNA |
| Rajaraman and Antani (2020) | Performed tri-level classification and studied task adaptation | Yes | RSNA pneumonia, PedPneumonia, Indiana, Shenzhen |
| Rajpurkar et al. (2020) | Developed a web-based system for TB affected HIV patients | Yes | CheXpert private dataset |
| Zhang et al. (2020) | Used deep model with Attention based CNN (CBAM) module | Yes | Shenzhen, Montgomery |
| Rahman M. et al. (2021) | Merged publicly available CXR dataset with XGBoost as classifier | Yes | Shenzhen, Montgomery |
| Owais et al. (2020) | Used a feature ensemble by combining low and high level features | Yes | Shenzhen, Montgomery |
| Das et al. (2021) | Modified a pre-trained InceptionV3 for TB classification | Yes | Shenzhen, Montgomery |
| Munadi et al. (2020) | Used enhancement techniques to improve deep classification | Yes | Shenzhen, Montgomery |
| Oloko-Oba and Viriri (2020) | Used deep learning-based pipeline for classification | Yes | Montgomery |
| Ul Abideen et al. (2020) | Proposed the Bayesian CNN to deal with uncertain TB and non-TB cases that have low discernibility. | Yes | Shenzhen, Montgomery |
| Hwang et al. (2016) | Proposed a modified AlexNet-based model for end-to-end training. Also performed cross-database evaluations. | Yes | Shenzhen, Montgomery |
| Gozes and Greenspan (2019) | Proposed MetaChexNet, trained on CXRs and metadata of gender, age and patient positioning. Later, finetuned the model for TB classification | Yes | ChestXray14, Shenzhen, Montgomery |
Pretraining (yes/no) refers to the use of weights of a deep model trained on ImageNet dataset. Private refers that the data used being in-house and is not released publicly.