Deployment of automated ultrasound venous gas emboli detection in real-world practice

Approved for Public Release It is well established that bubbles are a necessary, but not sufficient, condition for the development of decompression sickness (DCS). In scuba divers, these bubbles are assessed as circulating venous gas emboli (VGE) in blood post-dive using ultrasound, in either audio (precordial or subclavian Doppler) or imaging (trans-thoracic echocardiography) form. These recordings are evaluated in real-time and/or recorded for future (re-)analysis by expert human raters. Raters assign a semi-quantitative and non-linear severity grade ranging from no VGE to #white-out# where so many VGE are present that they partially mask normal background signals. Statistically, decompression sickness risk has been shown to increase with increasing VGE severity grade, although inter- and intra-personal differences have also been observed. Depending on the type of scale and recording used, these grading methods can be easier or more difficult and lengthy to learn, and inter-rater reliability can be low in less trained raters. Conversely, automating ultrasonic VGE assessment would allow for standardized, continuous, real-time decompression stress assessment. Our team has recently established that deep learning can be used to automate VGE assessment, with good performance in both Doppler and echocardiography. Our approach to model training focused on crowdsourced echocardiography spatio-temporal labels and Doppler synthetic data for transfer learning. In this project, we propose to refine and deploy our systems for Doppler in real-world settings, where we will also assess the real-time performance of our echocardiography final model, and continue building our repository of previously-acquired real Doppler dive data. To do so, we will continue working with our established team of collaborators at Duke and DAN as we devise the design for the integrated system to be tested in both chamber and field human dives, as well as UCSD for the repository and Doppler performance assessment. Our specific aims are as follows: (1) Refine and deploy the deep learning model for VGE grading in Doppler ultrasound; (2) Evaluate the real-time performance of the deep learning system for VGE detection in echocardiography in the field; and (3) Continue building the repository of previously acquired Doppler ultrasound data from historic experiments. Computer-automated VGE analysis in Doppler and echocardiography is important as it pertains to DCS risk mitigation. Decompression algorithms to predict DCS incidence are validated against observed DCS from large databases of exposures, but VGE are often recorded and used as a secondary endpoint. As such, the ability to accurately label such data in real-time can save analysis time and allow for more data to be taken into account for modeling and validation purposes. Building a repository of previously acquired Doppler data is also relevant to Navy capabilities. This large bank of data can be used to refine the automated algorithm for VGE assessment. Importantly, it will also safeguard historic experimental data that would be very difficult to reproduce nowadays and increase statistical power beyond what was previously feasible with individual studies, opening the door to future reanalysis of integrated data for possible new insights into DCS.