Machine Learning-Based Optimization of tFUS Transducer Positioning for Targeted Visual Cortex Neuromodulation

Abstract

Transcranial focused ultrasound (tFUS) is a promising non-invasive neuromodulation technique for stimulating the visual cortex and inducing visual perception through focused ultrasound waves. Accurate stimulation targeting remains challenging due to computationally intensive modeling of ultrasound wave propagation through heterogeneous media, particularly the skull. This study addresses these challenges by comparing multiple machine learning models - Neural Network (NN), Random Forest (RF), Linear Regression (LR), and Support Vector Regression (SVR) - for optimizing the positioning and focal length adjustment of tFUS transducers for targeted visual cortex stimulation. The approach leverages data from tFUS acoustic simulations with CT-derived skull models from 32 subjects and incorporates the proposed SphereFit model, which accounts for skull curvature in transducer positioning. The models were trained on features including the SphereFit-Derived skull center and radius, along with target focal coordinates, and were designed to predict the optimal transducer configurations: azimuth, elevation, and their corresponding focal points. The Random Forest model demonstrated superior performance among all tested approaches, achieving the lowest average focal point error of 2.76 ± 0.57 mm, suggesting the potential for maintaining precision in visual cortex targeting. This comparative study demonstrates the efficacy of machine learning approaches in optimizing tFUS stimulation, significantly reducing computational demands while maintaining high targeting accuracy.Clinical Relevance - The proposed method establishes the potential of machine learning-enhanced tFUS transducer positioning to improve non-invasive neuromodulation, enabling more precise and efficient targeting of the visual cortex for potential applications in vision restoration and neurological rehabilitation.