Diagnostic Accuracy of Artificial Intelligence for Pneumonia Detection Using Lung Ultrasound Images: A Systematic Review

ABSTRACT

Pneumonia continues to impose a substantial burden of morbidity and mortality globally. Lung ultrasound (LUS) has attracted growing clinical interest as a safe, portable, and radiation‐free imaging modality; however, variability in image interpretation is shaped by operator training. Artificial intelligence (AI) has been proposed as a means of standardising and enhancing the consistency of LUS interpretation in pneumonia diagnosis. This systematic review was conducted to evaluate the diagnostic performance of AI‐assisted LUS image interpretation for pneumonia detection across paediatric and adult populations, and to characterize the methodological heterogeneity of AI approaches represented in the current literature. A structured literature search was performed in accordance with PRISMA 2020 guidelines across PubMed, Google Scholar, Web of Science, Scopus, Elsevier, ResearchGate, and the Cochrane Library, covering publications from 2018 to 2025. Eligible studies were original research articles applying machine learning or deep learning algorithms to LUS image interpretation for pneumonia diagnosis, with at least one quantitative diagnostic performance metric reported. Ten studies satisfying the eligibility criteria were included, collectively encompassing 15,868 participants across paediatric, adult, and mixed clinical populations. AI architectures ranged from artificial neural networks and convolutional neural networks to hybrid frameworks, three‐dimensional video‐based models, and segmentation‐driven systems. Sensitivity ranged from 88% to 93%, specificity from 80% to 100%, and accuracy from 86% to 94.5% across the included studies. AI‐assisted interpretation of LUS images demonstrates strong diagnostic performance and holds promise as a tool for reducing inter‐observer variability in pneumonia detection.