Improved Nondestructive Ultrasound Molecular Imaging With Lightweight Convolutional Neural Network

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Ultrasound molecular imaging (USMI) uses targeted microbubbles (MBs) to detect disease biomarkers.
• Current state-of-the-art USMI, differential targeted enhancement (DTE), requires destructive pulses, limiting real-time clinical use.

Purpose of the Study:

• To develop and validate a neural network-based nondestructive USMI technique.
• To enable real-time, free-hand molecular imaging for clinical applications.

Main Methods:

• Developed a neural network for nondestructive USMI, trained on extensive data (15,350 patches) using augmentation strategies.
• Validated in vivo using a transgenic mouse model of spontaneous breast cancer.
• Compared performance against destructive differential targeted enhancement (DTE) and contrast-enhanced ultrasound (CEUS) reference images.

Main Results:

• Nondestructive USMI demonstrated high correlation (0.954) with DTE and superior AUC (0.954) compared to DTE (0.845) against a CEUS reference.
• Achieved a continuous Dice coefficient of 0.863 for molecular signal coverage.
• Outperformed DTE in AUC (0.953 vs. 0.892) under transducer motion, indicating potential for free-hand imaging.

Conclusions:

• Nondestructive USMI shows comparable performance to DTE under stationary conditions.
• This novel approach offers superior performance to DTE during transducer motion, highlighting its clinical imaging potential.
• The developed technique paves the way for real-time, non-destructive molecular imaging in clinical settings.