AI Empowers Nonexperts in Lung Ultrasound Imaging

This study explored whether artificial intelligence (AI) could help healthcare professionals who are not experts in lung ultrasound (LUS) capture high-quality images. The objective was to determine if AI guidance would allow trained healthcare professionals (THCPs) to produce diagnostic-quality images comparable to those obtained by LUS experts.

Lung ultrasound is essential for diagnosing conditions such as pneumonia and pulmonary edema, especially in patients with breathing issues. It is portable, affordable, and radiation-free. However, proper use requires skill. AI guidance for non-experts could broaden access to quality LUS in regions without skilled personnel, improving care and diagnostics in underserved areas.

Conducted across four clinical sites from July to December 2023, the study involved adults with shortness of breath who underwent two LUS exams: one by a THCP using AI and another by an LUS expert without AI. The AI, employing deep learning algorithms, guided image acquisition and automatically saved high-quality images using its “Autocapture” feature. A panel of expert readers, independent of those acquiring clips and unaware of the AI's involvement, assessed the image quality.

The study found that 98.3% of AI-assisted LUS exams by THCPs met diagnostic quality standards, comparable to those by LUS experts. This suggests AI can help non-experts produce expert-level images. The results were consistent across different patient demographics and a range of body mass indexes (BMIs).

AI guidance could democratize access to LUS, enabling non-expert healthcare workers to perform quality exams. This is particularly beneficial in resource-limited settings where expert LUS practitioners are scarce, enhancing diagnostic imaging capabilities across various healthcare environments.

Future research will focus on integrating AI for both image acquisition and interpretation, such as detecting lung conditions such as pleural effusions. Researchers plan to validate these systems in real-world clinical settings to assess their effectiveness and usability.

A minimum sample size of 130 patients was required to achieve 95% statistical power, and the analysis included an intention-to-treat sample of 176 subjects. The study found that 98.3% of AI-assisted LUS exams met diagnostic quality standards (95% CI, 95.1%-99.4%), with no significant difference compared to expert-acquired studies (difference, 1.7%; 95% CI, −1.6% to 5.0%; P = .31). In zone 6, THCPs outperformed experts (90.9% vs. 77.3%; difference, 13.6%; 95% CI, 6.1%-21.1%; P < .001). The mean examination time with AI was 16.5 minutes.

This research was funded by Caption Health (Now GE Healthcare) via grant INV-006401 from the Bill and Melinda Gates Foundation, and also by Yale University.