Convolutional neural network-based single-shot speckle tracking for x-ray phase-contrast imaging.

Area of Science:

• Medical Imaging
• Computational Imaging
• Biophysics

Background:

• X-ray phase-contrast imaging (XPCi) provides superior sensitivity for soft tissues but faces clinical barriers due to high coherence needs and costly optics.
• Speckle-based XPCi offers a cost-effective alternative, but its effectiveness hinges on precise tracking of speckle modulations caused by the sample.

Approach:

• Developed a convolutional neural network (CNN) to precisely determine sub-pixel displacement fields from reference and sample speckle images.
• Generated synthetic speckle patterns using wave-optical simulations, then introduced random deformations and attenuations for robust training and testing datasets.
• Validated the CNN against traditional speckle tracking methods, including zero-normalized cross-correlation and unified modulated pattern analysis.

Key Points:

• The CNN achieved 1.7x better accuracy, 2.6x reduced bias, and 2.3x improved spatial resolution compared to conventional methods.
• Demonstrated enhanced robustness to noise, independence from window size, and improved computational efficiency.
• Successfully validated the CNN approach using a simulated geometric phantom.

Conclusions:

• Introduced a novel CNN-based speckle tracking method that significantly enhances performance and robustness for speckle-based XPCi.
• This advanced tracking technique overcomes limitations of current methods, broadening the clinical applicability of affordable phase-contrast imaging.
• The CNN approach represents a significant step towards wider clinical adoption of XPCi for sensitive imaging of weakly-attenuating materials.