Challenges of Image-Based Diagnosis of Respiratory Diseases with Artificial Intelligence: A Systematic Review and Meta-Analysis

This article aims to synthesize the diagnostic accuracy of artificial intelligence (AI) for CT-based diagnosis of major respiratory diseases (COVID-19, tuberculosis [TB], COPD/ILA, lung nodules/cancer) between 2020 and 2025, and to identify barriers to clinical adoption spanning data standardization, interpretability, workflow integration, and radiation protection. Following PRISMA 2020, we searched PubMed/MEDLINE, Scopus, Web of Science, IEEE Xplore, and screened preprints (January 2020 to September 2025). Eligible human studies reported the diagnostic performance of AI (ML/DL/CNN/CAD) using CT. Primary outcomes were sensitivity, specificity, and AUC; secondary outcomes included CT dose metrics, explainability, and workflow effects. Risk of bias was assessed using QUADAS-2/PROBAST-AI; reporting quality was assessed using CLAIM. Bivariate random-effects meta-analysis yielded pooled estimates with HSROC; heterogeneity (τ 2, I 2) and publication bias (Deeks) were assessed. Certainty was graded using GRADE for DTA. Thirty-nine studies met criteria (predominantly CT; CNN-based). Pooled sensitivity/specificity were as follows: TB 0.895/0.935, COVID-19 0.872/0.914, nodules/cancer 0.886/0.869, COPD/ILA 0.856/0.844; heterogeneity was extreme (I 2 ≈ 99–100%). PROBAST-AI indicated highest concerns in analysis and predictors; CLAIM adherence was uneven (external validation: 33%, prospective evaluation: 12%). AI reduced reporting time by ∼20 to 40% and supported low-dose CT with ∼25 to 40% CTDIvol reductions while maintaining sensitivity. Deeks' plots suggested modest asymmetry. GRADE certainty was moderate (COVID-19, nodules/cancer), low–moderate (COPD), and low (TB). AI demonstrates promising diagnostic performance across respiratory CT tasks but faces generalizability, bias, and reporting gaps. Prospective multicenter validation, standardized protocols and dose reporting, calibrated/transparent models with quantitatively validated explainability, and assistive workflow deployment are essential for safe, reliable clinical adoption.