Artificial Intelligence Enables Nonexperts to Automatically Capture Lung Ultrasound Clips Containing B-Lines.
Authors
Affiliations (12)
Affiliations (12)
- Department of Emergency Medicine, Yale University School of Medicine, New Haven, CT, USA. Electronic address: [email protected].
- Department of Emergency Medicine, Feinberg School of Medicine, Northwestern University, Chicago, IL, USA.
- Department of Medicine, Feinberg School of Medicine, Northwestern University, Chicago, IL, USA.
- Cone Health, LeBauer Pulmonary and Critical Care, Greensboro, NC, USA; Department of Internal Medicine, University of North Carolina, Chapel Hill, NC, USA.
- Department of Emergency Medicine, Massachusetts General Hospital, Boston, MA, USA.
- Department of Emergency Medicine, Yale University School of Medicine, New Haven, CT, USA.
- Sidney Kimmel Medical College, Thomas Jefferson University, Philadelphia, PA, USA; Christiana Care Health System, Newark, DE, USA.
- Department of Emergency Medicine, University of Cincinnati, Cincinnati, OH, USA.
- Department of Surgical, Medical and Molecular Pathology and Critical Care Medicine, University of Pisa, Pisa, Italy.
- Caption Health / GE Healthcare, Chicago, IL, USA.
- Department of Emergency Medicine, Rush University Medical Center, Chicago, IL, USA.
- Cone Health, Greensboro, NC, USA.
Abstract
B-line artifacts on lung ultrasound (LUS) are the sonographic sign of partial deaeration of the lung, associated with conditions like acute pulmonary edema or pneumonia. While artificial intelligence (AI) has demonstrated promise in assisting with identification of B-lines, there remains a critical need for seamless integration of guidance, pathology identification, and auto-capture of pathological clips. This study evaluated a deep-learning algorithm for B-line annotation and auto-capture. This was a preplanned secondary analysis of a larger prospective multicenter validation trial of adult participants with shortness of breath, who underwent two ultrasound examinations following an 8-zone LUS protocol: one performed by a trained healthcare professional (THCP) using Lung Guidance AI, and the other by a fellowship-trained LUS expert without AI assistance. Five blinded expert LUS readers provided remote review and ground truth validation. B-line detection and auto-capture were analyzed using balanced accuracy (BA) and positive predictive value (PPV). The performance of the B-line severity function was assessed by scaled Gwet's Agreement Coefficient (AC1/AC2) between the B-line tool and the ground truth for B-line severity scores. B-line detection and auto-capture accuracy was strong (BA = 78%; 95% CI: 74.2%- 80.8%; PPV = 91%; 95% CI: 80.4%-96.1%), and the severity function had strong agreement with the ground truth ratings (Gwet's AC1/AC2 = 96%; 95% CI: 94.0%-97.9%). AI-assisted B-line detection, auto-capture, and severity assessment in LUS can achieve high accuracy and agreement with expert interpretations, potentially improving standardization and efficiency in clinical practice across various patient populations and operator skill levels.