New lung ultrasound system for rapid triage of pulmonary disease without a radiologist or sonographer

Abstract

BACKGROUND: Most people in the world lack access to medical imaging including for assessment of pulmonary disease. We sought to improve access to pulmonary imaging by developing a rapid automated system for triage of pulmonary disease using lung ultrasound requiring neither a radiologist nor experienced sonographer utilizing volume sweep imaging (VSI) and artificial intelligence (AI). METHODS: We conducted a retrospective study of lung ultrasound VSI data collected from May 2019 to January 2020. AI analysis utilizing a convolutional neural network and random forest-based machine learning classifier was performed on 70 normal lung ultrasound VSI video clips and 49 abnormal lung ultrasound VSI video clips obtained by individuals without prior ultrasound experience. The accuracy of the AI was assessed for the ability to distinguish between normal and abnormal lung ultrasound video clips. RESULTS: Among test VSI clips (n = 36 clips), AI achieved 91.7% accuracy, 85.7% sensitivity, 95.5% specificity, and an F1 score of 0.89 for an abnormal lung ultrasound VSI clip. Among test subjects (n = 20) from which these clips were obtained, 90.0% accuracy, 87.5% sensitivity, 91.7% specificity, and an F1 score of 0.88 were achieved. CONCLUSIONS: Lung ultrasound VSI integrated with AI shows potential to provide preliminary triage of pulmonary disease allowing a system for rapid automatic triage of pulmonary disease requiring neither a radiologist nor sonographer.

Ähnliche Arbeiten