Automated analysis of ultrasound-activated microbubble interactions using machine learning: SO3 P3preliminary results

Shelled microbubbles are recognized for their contrast enhancing properties in the context of diagnostic ultrasound as well as their emerging therapeutic potential. Whilst theoretical understanding, validated via experimental observations, have significantly improved the knowledge base in inter-bubble interactions over recent years, the required image analysis procedure has remained a largely manual process involving frame-by-frame measurement methodologies. Here, we present an automated approach that attempts to address that slow and error-prone process. Our strategy employs data processing and machine-learning (ML) based detection to reliably identify in-scene microbubbles. Our preliminary results are presented, duly validated with a selection of high-speed footage from optically controlled microbubble interactions. In the present incarnation of our ML routine, bubble radii and centroid displacements may be tracked automatically and thus lead to accurate automated assessment of Bjerknes forces and microbubble compressibility for example. Comparative errors calculated from the ratio of machine-learning inferences to expertly assessed manual measurements shows that the ML analysis is correct to within 0.44% on average, with a maximal error of 2.28%. Whilst the speed advantage to our ML approach is clear, and accuracy levels certainly appear to be acceptable, some restrictions to the automated approach are also highlighted.