Frances M Russell1, Robert R Ehrman2, Allen Barton3, Elisa Sarmiento4, Jakob E Ottenhoff2, Benjamin K Nti4. 1. Department of Emergency Medicine, Indiana University School of Medicine, 720 Eskenazi Ave, FOB 3rd Floor, Indianapolis, IN, 46202, USA. framruss@iu.edu. 2. Department of Emergency Medicine, Wayne State University School of Medicine, 4021 St Antoine Ave, Suite 6G, Detroit, MI, 48201, USA. 3. Boone County Emergency Physicians, Zionsville, IN, 46077, USA. 4. Department of Emergency Medicine, Indiana University School of Medicine, 720 Eskenazi Ave, FOB 3rd Floor, Indianapolis, IN, 46202, USA.
Abstract
BACKGROUND: The goal of this study was to assess the ability of machine artificial intelligence (AI) to quantitatively assess lung ultrasound (LUS) B-line presence using images obtained by learners novice to LUS in patients with acute heart failure (AHF), compared to expert interpretation. METHODS: This was a prospective, multicenter observational study conducted at two urban academic institutions. Learners novice to LUS completed a 30-min training session on lung image acquisition which included lecture and hands-on patient scanning. Learners independently acquired images on patients with suspected AHF. Automatic B-line quantification was obtained offline after completion of the study. Machine AI counted the maximum number of B-lines visualized during a clip. The criterion standard for B-line counts was semi-quantitative analysis by a blinded point-of-care LUS expert reviewer. Image quality was blindly determined by an expert reviewer. A second expert reviewer blindly determined B-line counts and image quality. Intraclass correlation was used to determine agreement between machine AI and expert, and expert to expert. RESULTS: Fifty-one novice learners completed 87 scans on 29 patients. We analyzed data from 611 lung zones. The overall intraclass correlation for agreement between novice learner images post-processed with AI technology and expert review was 0.56 (confidence interval [CI] 0.51-0.62), and 0.82 (CI 0.73-0.91) between experts. Median image quality was 4 (on a 5-point scale), and correlation between experts for quality assessment was 0.65 (CI 0.48-0.82). CONCLUSION: After a short training session, novice learners were able to obtain high-quality images. When the AI deep learning algorithm was applied to those images, it quantified B-lines with moderate-to-fair correlation as compared to semi-quantitative analysis by expert review. This data shows promise, but further development is needed before widespread clinical use.
BACKGROUND: The goal of this study was to assess the ability of machine artificial intelligence (AI) to quantitatively assess lung ultrasound (LUS) B-line presence using images obtained by learners novice to LUS in patients with acute heart failure (AHF), compared to expert interpretation. METHODS: This was a prospective, multicenter observational study conducted at two urban academic institutions. Learners novice to LUS completed a 30-min training session on lung image acquisition which included lecture and hands-on patient scanning. Learners independently acquired images on patients with suspected AHF. Automatic B-line quantification was obtained offline after completion of the study. Machine AI counted the maximum number of B-lines visualized during a clip. The criterion standard for B-line counts was semi-quantitative analysis by a blinded point-of-care LUS expert reviewer. Image quality was blindly determined by an expert reviewer. A second expert reviewer blindly determined B-line counts and image quality. Intraclass correlation was used to determine agreement between machine AI and expert, and expert to expert. RESULTS: Fifty-one novice learners completed 87 scans on 29 patients. We analyzed data from 611 lung zones. The overall intraclass correlation for agreement between novice learner images post-processed with AI technology and expert review was 0.56 (confidence interval [CI] 0.51-0.62), and 0.82 (CI 0.73-0.91) between experts. Median image quality was 4 (on a 5-point scale), and correlation between experts for quality assessment was 0.65 (CI 0.48-0.82). CONCLUSION: After a short training session, novice learners were able to obtain high-quality images. When the AI deep learning algorithm was applied to those images, it quantified B-lines with moderate-to-fair correlation as compared to semi-quantitative analysis by expert review. This data shows promise, but further development is needed before widespread clinical use.
Authors: Francesco Corradi; Claudia Brusasco; Alessandro Garlaschi; Francesco Paparo; Lorenzo Ball; Gregorio Santori; Paolo Pelosi; Fiorella Altomonte; Antonella Vezzani; Vito Brusasco Journal: Biomed Res Int Date: 2015-02-25 Impact factor: 3.411