Neural Operators for Optimizing Focused Ultrasound Therapy

Combining physics and AI for rapid and precise predictions of optimal therapeutic site locations.

This research focuses on developing a model to enable personalized treatment plans for spinal cord injury. Understanding where to place a focused ultrasound transducer in surgery to optimize therapeutic effect requires computational simulations. Numerical solvers take too long to generate the necessary pressure maps, taking up to 5 hours to provide accurate results. The time and cost-intensive nature of this approach is unfeasible for intraoperative use-cases. My approach uses a physics-informed deep operator network trained on simulated pressure maps in the spinal cord to predict the output pressure distribution for a given patient in a matter of seconds. Regularized by physical constraints, this novel architecture learns the mappings between the patient-specific anatomy and the solution for the governing wave equation (pressure distribution) to approximate the overarching operator. This presents a paradigm-shifting solution to personalizing spinal cord care, overcoming the computational burden of running several expensive simulations.

So far, the designed network has been able to achieve 2% error in the test set using a segmented mask of the ultrasound spinal cord image! This work is under preparation and will be available online soon!