DefCor-Net: Physics-aware ultrasound deformation correction.

Area of Science:

• Medical imaging
• Biomedical engineering
• Artificial intelligence in healthcare

Background:

• Ultrasound (US) image acquisition often results in deformed anatomical images, challenging accurate diagnosis.
• Accurate anatomical reconstruction is vital for computer-assisted diagnosis systems.
• Existing methods struggle with precise deformation correction in real-time US imaging.

Purpose of the Study:

• To present a novel physics-aware deep neural network, DefCor-Net, for accurate deformation correction in ultrasound images.
• To enable anatomy-aware correction by incorporating biomedical knowledge and real-time tissue property estimation.
• To improve the recovery of original geometry in deformed ultrasound images.

Main Methods:

• Developed a coarse-to-fine, multi-scale deep neural network (DefCor-Net) for deformation correction.
• Integrated biomedical knowledge by estimating pixel-wise tissue stiffness online using a U-shaped feature extractor.
• Computed the deformation field using polynomial regression based on measured probe force.

Main Results:

• DefCor-Net significantly enhanced deformation correction accuracy, improving the Dice Coefficient from 14.3±20.9 to 82.6±12.1 (at 6N force).
• The model demonstrated effective recovery of original anatomical geometry from deformed ultrasound images.
• Validation was performed on ultrasound images from volunteer forearms and upper arms.

Conclusions:

• The proposed DefCor-Net offers a powerful, learning-based approach for anatomy-aware deformation correction in ultrasound imaging.
• Real-time estimation of tissue properties enables improved geometric accuracy, crucial for diagnostic consistency.
• This method holds potential for advancing computer-assisted diagnosis through enhanced image fidelity.