Accuracy of deep learning for automated detection of pneumonia using chest X-Ray images: A systematic review and meta-analysis

Abstract

Background: Recently, deep learning (DL) algorithms have received widespread popularity in various medical diagnostics. This study aimed to evaluate the diagnostic performance of DL models in the detection and classifying of pneumonia using chest X-ray (CXR) images.

Methods: PubMed, Embase, Scopus, Web of Science, and Google Scholar were searched in order to retrieve all studies that implemented a DL algorithm for discriminating pneumonia patients from healthy controls using CXR images. We used bivariate linear mixed models to pool diagnostic estimates including sensitivity (SE), specificity (SP), positive likelihood ratio (PLR), negative likelihood ratio (NLR), and diagnostic odds ratio (DOR). Also, the area under receiver operating characteristics curves (AUC) of the included studies was used to estimate the diagnostic value.

Results: The pooled SE, SP, PLR, NLR, DOR and AUC for DL in discriminating pneumonia CXRs from controls were 0.98 (95% confidence interval (CI): 0.96-0.99), 0.94 (95% CI: 0.90-0.96), 15.35 (95% CI: 10.04-23.48), 0.02 (95% CI: 0.01-0.04), 718.13 (95% CI: 288.45-1787.93), and 0.99 (95% CI: 0.98-100), respectively. The pooled SE, SP, PLR, NLR, DOR and AUC for DL in discriminating bacterial from viral pneumonia using CXR radiographs were 0.89 (95% CI: 0.79-0.94), 0.89 (95% CI: 0.78-0.95), 8.34 (95% CI: 3.75-18.55), 0.13 (95% CI: 0.06-0.26), 66.14 (95% CI: 17.34-252.37), and 0.95 (0.93-0.97).

Conclusion: DL indicated high accuracy performance in classifying pneumonia from normal CXR radiographs and also in distinguishing bacterial from viral pneumonia. However, major methodological concerns should be addressed in future studies for translating to the clinic.

Keywords: Artificial intelligence; CXR; Deep learning; Meta-analysis; Pneumonia.