Artificial Intelligence-Based Automated Analysis for Pleural Effusion Detection on Thoracic Ultrasound: A Systematic Review

Background: Pleural effusion (PE) is a common condition where accurate detection is essential for management. Thoracic ultrasound (TUS) is the first-line modality owing to safety, portability, and high sensitivity, but accuracy is operator-dependent. Artificial intelligence (AI)-based automated analysis has been explored as an adjunct, with early evidence suggesting potential to reduce variability and standardise interpretation. This review evaluates the diagnostic accuracy of AI-assisted TUS for PE detection. Methods: This review was registered with PROSPERO (CRD420251128416) and followed PRISMA guidelines. MEDLINE, Scopus, Google Scholar, IEEE Xplore, Cochrane CENTRAL, and ClinicalTrials.gov were searched through 20 August 2025 for studies assessing AI-based TUS analysis for PE. Eligible studies required recognised reference standards (expert interpretation or chest CT). Risk of bias was assessed with QUADAS-2, and certainty with GRADE. Owing to heterogeneity, structured narrative synthesis was performed instead of meta-analysis. Results: Five studies (7565 patients) published between 2021-2025 were included. All used convolutional neural networks with varied architectures (ResNet, EfficientNet, U-net). Sensitivity ranged 70.6-100%, specificity 67-100%, and AUC 0.77-0.99. Performance was reduced for small, trace, or complex effusions and in critically ill patients. External validation showed attenuation compared with internal testing. All studies had high risk of bias in patient selection and index test conduct, reflecting retrospective designs and inadequate dataset separation. Conclusions: AI-assisted TUS shows promising diagnostic performance for PE detection in curated datasets; however, evidence is inconsistent and limited by key methodological weaknesses. Overall certainty is low-to-moderate, constrained by retrospective designs, limited dataset separation, and scarce external validation. Current evidence is insufficient to support routine clinical use. Robust prospective multicentre studies with rigorous independent validation and evaluation of clinically meaningful outcomes are essential before clinical implementation can be considered.