Wavelet-inspired diffusion model with near-field constraint for real-time echocardiography dehazing

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Echocardiography is crucial for cardiac diagnosis but suffers from near-field haze due to sound speed variations.
• Existing dehazing methods lack efficacy and efficiency due to simplified assumptions.
• This limits diagnostic accuracy and clinical utility.

Purpose of the Study:

• To develop a real-time echocardiography dehazing framework, EchoWDiff.
• To improve the reconstruction of cardiac anatomy and diagnostic accuracy.
• To address limitations of current dehazing techniques.

Main Methods:

• Proposed EchoWDiff, a wavelet-inspired diffusion model with near-field constraint.
• Introduced an adversarial pairing module for realistic clean-hazy pair generation.
• Integrated multi-scale frequency analysis and contrastive perceptual loss for enhanced dehazing.

Main Results:

• Achieved significant improvements in echocardiography dehazing metrics (FID, gCNR).
• Demonstrated superior performance in near-field left ventricle segmentation (Dice, ASD).
• Showcased real-time processing speed (27 FPS) with promising clinical applicability.

Conclusions:

• EchoWDiff effectively removes near-field haze in echocardiography.
• The framework enhances anatomical detail and diagnostic accuracy.
• The method offers a computationally efficient and clinically viable solution.