Machine learning-assisted closed-loop focused ultrasound for targeted nanoparticle delivery and liquid biopsy in the brain

Focused ultrasound (FUS) combined with microbubbles (MB) provide a unique physical method to reversibly increase the permeability of the BBB and improve drug delivery. Real-time methods to monitor and control the MB dynamics is increasingly important as safe and effective treatment of larger brain regions over multiple sessions are becoming clinical priority. However, for current control methods, safety remains a concern because it primarily relies on reacting to MB collapse (i.e., strong wideband emissions—WBE). We hypothesize that machine learning (ML) can assist controllers to predict the onset of strong WBE (>6 dB) and adjust the exposure before WBE occurs, thereby attaining safer BBB opening without compromising efficacy. To test our hypothesis, we trained a multi-layer perceptron (MLP) model using past BBB opening acoustic emission dataset in rodents (54 040 total data). Subsequently, we integrated MLP to closed-loop controller and assessed its ability to predict WBE in real-time during BBB opening. In our training/testing dataset, MLP model showed 95% accuracy of predicting WBE; its application to real-time controller showed more than 90% reduction in total WBE events compared to current state-of-art closed-loop controller without compromising safety (assessed using immunohistochemistry) while retaining high BBB permeability (measured using dynamic contrast enhanced MRI)