<![CDATA[Pneumonia and COVID-19 remain leading causes of universal morbidity and mortality, with timely and precise diagnosis essential for effective patient management. This systematic review and meta-analysis assessed the diagnostic accuracy of deep neural networks in detecting pneumonia and COVID-19 across main medical imaging modalities. Comprehensive searches of PubMed, Scopus, Web of Science, IEEE Xplore and Cochrane Library identified 80 eligible studies published between 2017 and 2025. Included studies used chest X-ray (CXR), computed tomography (CT) and lung ultrasound (LUS) images analyzed through convolutional neural networks, transformer-based and hybrid deep models. Pooled diagnostic performance was synthesized using a bivariate random-effects model and hierarchical summary receiver operating characteristic analysis. Overall pooled sensitivity and specificity were 0.88 (95% CI: 0.84-0.91) and 0.90 (95% CI: 0.86-0.92), respectively, with an area under the curve of 0.93, indicating high discriminative capability. Subgroup analyses revealed CT-based models outperformed CXR and LUS, while transformer architectures marginally exceeded CNNs. In addition, external validation studies steadily reported lower accuracy than internal validations, reflecting limited model generalizability. Risk of bias assessment using QUADAS-2 emphasized concerns related to patient selection, data leakage and non-standardized reference criteria. Despite moderate heterogeneity (I² = 39-52%) and potential publication bias, findings confirm the substantial potential of DNNs as decision-support tools for fast, scalable and reliable respiratory disease diagnosis. However, broader clinical adoption demands multicenter validation, transparency and adherence to ethical AI standards. This study provides evidence-based insights into the current performance and translational readiness of AI-driven diagnostic imaging for pneumonia and COVID-19.]]>
