Towards Personalized Decompression Medicine: Applied Deep-Learning for Automated Doppler and Echocardiogram Analysis

It is widely accepted that bubbles are a necessary but insufficient condition for the development of decompression sickness (DCS). In decompression research, information about bubble presence is gathered using ultrasound acquisitions, using Doppler audio where bubble flow is detected, and B-mode echocardiography, where bubbles or venous gas emboli (VGE) can be visualized passing through the heart chambers. It has been well established that there is a positive association between VGE detected with ultrasound and the risk of DCS. As such, ultrasound measurement post-dive has become a staple tool for decompression researchers. Towards understanding subject-dependent variability to dives, there has been a greater push towards the use of continuous measurements post-dive, rather than at a single or few timepoints, providing a new avenue to study the effect of dive profiles and preconditioning treatments on susceptibility to DCS. However, the main limitation of collecting these large amounts of data lies within the analysis, which is time-consuming and requires skilled raters to manually assess. To allow for large-scale studies to be performed and provide decompression researchers with tools to accelerate their progress, robust and accurate automation systems need to be developed and evaluated. In this dissertation, we explore five new methods for fully automated analysis of both post-dive Doppler audio and B-mode echocardiograms. First, we develop and evaluate a method for automated heart rate estimation in post-dive Doppler audio, which could potentially be combined with VGE counting algorithms to approximate VGE grading. Subsequently, we establish a technique for synthetic Doppler audio data generation to supplement the real-world dataset available. Furthermore, using these two systems, alongside the field’s largest compiled real-world dataset, we train a novel deep learning framework for VGE classification, showing substantial agreement with human raters on both subclavian vein and precordial Doppler measurements that is on par with human inter-rater agreement for Doppler analysis. Fourth, we introduce and validate the use of deep learning object detectors for VGE detection in echocardiograms. Finally, we describe methods for automated late ventricular diastole frame selection and heart chamber selection and their integration with the previous VGE detection network for an end-to-end fully automated frame-based bubble counting system, demonstrating greater internal consistency and lower bias compared to human raters.