Respiratory Signal Estimation for Free-Hand 2D Ultrasound Image via Heterogeneous Alignment and Conditional...

Area of Science:

• Medical imaging
• Signal processing
• Artificial intelligence

Background:

• 2D ultrasound captures organ motion due to respiration, crucial for estimating respiratory signals.
• Free-hand ultrasound acquisition introduces spatial motion, compromising respiratory motion estimation accuracy.

Purpose of the Study:

• To propose an unsupervised method for accurate respiratory signal estimation in ultrasound.
• To address and mitigate spatial motion interference in free-hand ultrasound acquisitions.

Main Methods:

• A heterogeneous graph of ultrasound slices was constructed to differentiate organ respiratory motion and probe spatial motion.
• Hierarchical attention aggregation was used to independently learn respiratory and spatial correlations.
• Conditional generative learning mapped ultrasound images into a unified respiratory motion feature space for signal estimation.

Main Results:

• The proposed method achieved over 93% correlation with ground truth in various free-hand acquisition modes, reaching up to 97%.
• The model estimates respiratory signals for a single slice in approximately 2 milliseconds.
• Demonstrated superior accuracy and robustness compared to existing methods.

Conclusions:

• The unsupervised heterogeneous information alignment method effectively estimates respiratory signals from free-hand ultrasound, overcoming spatial motion artifacts.
• This technique enhances accuracy and robustness, offering a foundation for real-time 4D respiratory ultrasound imaging without fixed position constraints.