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

Abstract

Background

In ultrasound imaging, the generated images involve speckle noise owing to the mechanism underlying image generation. Speckle noise directly affects image analysis, necessitating its effective suppression.

Purpose

Ultrasound image denoising offers limited performance and causes structural information loss. To address these challenges and improve ultrasound image quality, we develop a new denoising method based on the diffusion model (DM).

Methods

This exploratory study proposes a DM-based denoising method, namely adversarial DM with feature extraction network (ADM-ExNet) to investigate the potential of combining diffusion models and generative adversarial Networks (GANs) for ultrasound image denoising. Specifically, we replace the reverse process of the DM with a GAN and modify the generator and discriminator as a U-Net structure. Simultaneously, a structural feature extraction network is incorporated into the model to construct a loss function, which offers enhanced detail retention. The noise levels ($\sigma =10,15,20$) were simulated by adding Gaussian noise to the original ultrasound images, where $\sigma$ controls the intensity of the noise. We employed three public datasets, HC18, CAMUS, and Ultrasound Nerve, which involve the ultrasound images of the fetal head circumference, heart, and nerves, respectively. Each image was adjusted to $256\times 256$ pixels, and the training set and the validation set were divided by 9:1. The mean square error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) were employed as primary evaluation metrics. To rigorously validate the statistical significance of performance differences, we further applied false discovery rate (FDR) correction for hypothesis testing and calculated Cohen's d effect sizes to quantify the magnitude of improvements against baselines. ADM-ExNet was compared with three traditional filtering methods and four deep learning methods with the U-Net structure.

Results

The proposed ADM-ExNet significantly enhances denoising performance across all datasets, with PSNR improvements exceeding 12 dB over noisy baselines and MSE reductions of over 90%. Notably, ADM-ExNet achieves high SSIM values (e.g., 0.941 at $\sigma =10$ on HC18 vs. 0.369 for noisy images), demonstrating superior structural preservation without excessive smoothing. Statistical significance (FDR-adjusted ) and Cohen's d effect sizes (up to d = 3.8 on CAMUS at $\sigma =20$) confirm its robustness, outperforming traditional methods and deep learning competitors in both visual quality and quantitative metrics (PSNR, SSIM) across noise levels. This balance of detail retention and noise suppression highlights the exploratory potential of ADM-ExNet.

Conclusions

The proposed method improves the quality of ultrasound images with various structural features, effectively reducing noise while retaining details.