19 Automated Quantification Of B-Lines in Lung Ultrasound On COVID-19 Patients

Baloescu, C.; Chen, A.; Raju, B.; Evans, N.; Moore, C. L.

Study Objectives: Point-of-care ultrasound (POCUS) has become an important tool in the global response to COVID-19, supporting screening, diagnosis and management. Lung features described in COVID-19 include B-lines (ring-down artifacts appearing in the presence of interstitial lung fluid), thickened and irregular pleural line, subpleural consolidations and effusions. Computer-aided interpretation can be incorporated into POCUS platforms to provide objective data and improve interpretations by novices. We sought to test a commercially available B-Lines counting feature previously developed using non-COVID data, on patients suspected of COVID-19. This first step would allow expansion to automated B-line scoring and further lung feature detection, to create an intelligent POCUS system with comprehensive set of lung ultrasound features for COVID-19 pneumonia. Methods: This was a prospective observational study at a single academic medical center. Subjects presenting to the emergency department with shortness of breath and suspected COVID-19 were enrolled. The Philips Lumify TM ultrasound system with sector or linear transducer was used to obtain 6-second clips of 14 lung zones (upper and lower, right and left anterior, lateral and posterior). Right and left anterior upper zone clips were also obtained with a second probe type. Repeat examinations with data collection were performed on days 3, 5, 7 or 12+/- 1 day for admitted patients. All clips with 2 or more B-lines were included (N=80), as well as a random selection of 70 clips with 1 or fewer B-lines. B-line count for inclusion was based on visual rating by two researchers with POCUS training. A POCUS fellowship trained emergency physician visually assessed each clip frame and counted the maximum number of B-lines per clip. This was compared to automatic counts by the commercially available Lumify TM Lung B-lines Quantification software by intraclass correlation coefficient (ICC) and Cohen’s weighted kappa. Results: Of the 899 total clips,150 clips from 30 unique subjects and 44 overall exams were used for analysis, with 100 clips from patients with confirmed COVID by PCR. The average maximum B-line count by algorithm was 1.52 +/- 1.24, and that by expert was 1.60 +/- 1.35 (ns). The ICC between algorithm and expert was 0.87 (95% CI 0.83-0.91), with a weighted kappa of 0.64 (95% CI 0.48-0.81), indicating substantial agreement. Average of maximum B-line counts, ICC and weighted kappa between algorithm and expert were comparable for COVID+ and COVID- subgroups as well as between transducer types. For COVID + subgroup, the average of maximum B-line counts was 1.73 +/- 1.28 for algorithm and 1.78 +/- 1.37 for expert, with weighted kappa 0.67 (95% CI 0.50-0.84), and ICC 0.87 (95% CI 0.83 to 0.91). Conclusion: An automated algorithm developed on non-COVID patients can accurately distinguish and quantify B-lines in clips from patients with COVID-19, with substantial agreement to expert visual rating. [Formula presented]