Table 1.

A literature review of research articles investigating the role of AI in detecting pathologies from chest radiographs.

Study	Size of Test/Validation Dataset	Problem Statement	Method	Results	Advantages	Limitations
Annarumma et al., 2019 [17]	3229 institutional adult chest radiographs	Developed and tested an AI model, based on deep CNNs, for automated triaging of adult chest radiographs based on the urgency of imaging appearances	Ensemble of two deep CNNs	Sensitivity of 71%, specificity of 95%, PPV of 73%, and NPV of 94%	The AI model was able to interpret and prioritize chest radiographs based on critical or urgent findings	False Negatives could result in misinterpretation of urgent cases as non-urgent, delaying timely clinical attention Same radiology label could correspond to different levels of urgency. The spectrum of urgency was not addressed in the study
Dunnmon et al., 2019 [18]	533 frontal chest radiographs	Assessed the ability of CNNs to enable automated binary classification of chest radiographs	Variety of classification CNNs	AUC of 0.96	Demonstrated the automated classification of chest radiographs as normal or abnormal	Only predicted the presence or absence of abnormality in the thoracic region Did not provide explainability
Nguyen et al., 2022 [19]	6285 frontal chest radiographs	Deployed and validated an AI-based system for detecting abnormalities on chest X-ray scans in real-world clinical settings	EfficientNet	F1 score of 0.653, accuracy of 79.6%, sensitivity of 68.6%, and specificity of 83.9%	Examined the AI performance on a clinical dataset different from the training dataset	Classified radiographs into normal or abnormal due to lack of detailed ground truth Did not check the effect of AI on radiologist diagnostic performance
Saleh et al., [20]	18,265 frontal-view chest X-ray images	Developed CNN-based DL models and compared their feasibility and performance to classify 14 chest pathologies found on chest X-rays	Variety of classification CNNs with DC-GANs	Accuracy of 67% and 62% for the best-performing model with and without augmentation, respectively	Used GAN-based techniques for data augmentation to address the lack of data for some pathologies	A different test set was used for the AI model with augmentation Test sets included images from the NIH database only
Hwang et al., [21]	1089 frontal chest X-ray images	Developed a deep learning–based algorithm that classified chest radiographs into normal and abnormal for various thoracic diseases	Variety of classification CNNs	AUC of 0.979, sensitivity of 0.979, and specificity of 0.880	AI model outperformed physicians, including thoracic radiologists. Radiologists aided with DLAD performed better than radiologists without the aid of DLAD	Validation was performed using experimentally designed data sets and included chest radiographs with only 1 target disease DLAD covered only 4 major thoracic disease categories