Real-world usability of AI-based lung ultrasound systems

Determine the real-world usability of artificial intelligence models for lung ultrasound analysis trained on existing datasets of poor quality, specifically assessing whether such models generalize sufficiently for clinical decision support in prospective clinical settings.

Background

Lung ultrasound is increasingly used to assess lung pathologies, and artificial intelligence offers promise to alleviate time- and expertise-intensive interpretation. However, many prior AI studies rely on heterogeneous and biased datasets, raising concerns about generalizability to routine clinical practice.

This paper introduces a prospective dataset (COVID-BLUeS) and evaluates seminal AI approaches, highlighting discrepancies between previously reported performances and results on standardized prospective data. The authors explicitly note uncertainty about the usability of AI systems trained on poor-quality existing datasets for real-world applications.