DecoBubbles.com: The first crowdsourcing initiative to expedite VGE counting in 2D-ultrasound recordings and provide labels for machine learning

Abstract

Introduction/Background Venous gas emboli (VGE) detected with echocardi­ography post-dive are often used as a marker of decompression stress. The very time-consuming analysis of the acquired videos can help elucidate inter- and intrasubject variability in VGE, and DCS susceptibility. This collaborative project aims to reduce the time needed to count and analyze cardiac ultrasound videos and help train an Al to complete the task. To accomplish this, an interactive web platform 'www.decobubbles.com' was developed that allows crowdsourcing the counting process to lay persons without previous expertise in ultrasound analysis. Materials and Methods Volunteers from around the world are encouraged to register on the website to help with this initiative. After registration the volunteer is prompted to watch an eight-minute training tutorial teaching them how to choose suitable frames in the videos, count VGE and submit their assessment. A volunteer is allowed to participate in the project once they have succeeded in completing a number of training videos in which they need to agree with the counts of professional graders. A database of 204 de-identified post-dive ultrasound videos of variable quality have been uploaded to the application. Results The application on www.decobubbles.com is active since February 2021 and at the beginning of March 2021 had registered 153 unique raters, 587 submitted ratings, and 34% of the 204 videos in the database were fully rated, which means that three independent raters agreed in their counting. Summary/Conclusion This collaborative platform will allow refinement of an Al to count VGE in post-dive echocardiography recordings and advance the knowledge about inter­and intrasubjectVGE variability in divers. Future implications include the possibility for researchers to use the platform to directly analyze their own datasets - with no obligation to add these to our database, although we welcome more videos from other sources and collaborators. This research is funded through the Divers Alert Network (#DAN-UNC-1) and the Office of Naval Research (ONR #N00014-20-1-2590).