An exploratory study on ultrasound image denoising using feature extraction and adversarial diffusion model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Speckle noise in ultrasound images degrades image quality and hinders analysis.
• Existing denoising methods often compromise structural information.

Purpose of the Study:

• To develop a novel diffusion model (DM) based denoising method for ultrasound images.
• To enhance ultrasound image quality by effectively suppressing noise and retaining structural details.

Main Methods:

• A diffusion model (DM) combined with a generative adversarial network (GAN) was developed, named ADM-ExNet.
• The model incorporates a U-Net structure for generator and discriminator and a feature extraction network for detail retention.
• Experiments were conducted on HC18, CAMUS, and Ultrasound Nerve datasets with simulated Gaussian noise.

Main Results:

• ADM-ExNet demonstrated significant denoising performance, with PSNR improvements over 12 dB and MSE reductions over 90%.
• The method achieved high structural similarity index (SSIM) values, indicating superior structural preservation.
• Statistical analysis confirmed ADM-ExNet's robustness and superiority over traditional and deep learning methods.

Conclusions:

• The proposed ADM-ExNet effectively reduces noise in ultrasound images while preserving essential details.
• This method shows promise for improving the quality and diagnostic utility of ultrasound imaging.