Accelerating Focused Ultrasound Modeling in Heterogeneous Spinal Cord Anatomy With Vision Transformer Operator...

Area of Science:

• Biomedical Engineering
• Acoustic Physics
• Artificial Intelligence

Background:

• Focused ultrasound therapy offers precise, submillimeter targeting for spinal cord injuries.
• Accurate prediction of ultrasound pressure fields is crucial for treatment efficacy but computationally intensive.
• Current methods are too slow for real-time surgical decision-making.

Purpose of the Study:

• To develop a rapid, data-driven framework for predicting focused ultrasound pressure fields in the spinal cord.
• To overcome the computational limitations of traditional acoustic simulations.
• To enable real-time, patient-specific therapeutic ultrasound imaging.

Main Methods:

• Developed a Vision Transformer-based Deep Operator Network (ViT-DeepONet).
• Benchmarked ViT-DeepONet against convolutional neural network (CNN)-DeepONet and a geometry-aware Fourier Neural Operator.
• Trained models on 8000 simulated pressure maps from porcine spinal cords.

Main Results:

• ViT-DeepONet achieved real-time predictions with 3.1% test loss on porcine data and 4.4% on human ultrasound images.
• Computation time was reduced over 91,000-fold compared to CPU-based solvers.
• ViT-DeepONet outperformed CNN-DeepONet and offered a favorable balance of accuracy and complexity.

Conclusions:

• ViT-DeepONet accurately predicts ultrasound pressure fields in complex acoustic environments.
• Operator learning accelerates acoustic modeling for therapeutic ultrasound.
• The framework facilitates optimized, patient-specific treatment planning.