Denoising Plane Wave Ultrasound Images Using Diffusion Probabilistic Models

TL;DR

The paper addresses denoising in high-frame-rate plane wave ultrasound by leveraging a diffusion probabilistic model tailored for RF beamformed data, moving from low-quality to high-quality images. It introduces a forward interpolation $x_t = X_0 + t (X_1 - X_0)$ and a diffusion-based reverse process that predicts a velocity $v_\theta$, enabling efficient denoising from single to multi-angle plane waves while preserving speckle. A key novelty is using natural image segmentation masks as intensity maps to diversify anatomical shapes and train on 400 simulated images, achieving strong generalization to phantom and in vivo data. Across simulated, phantom, and in vivo experiments, the method demonstrates improved image quality and outperforms several baselines on metrics such as CNR, gCNR, NRMSE, and SSIM, with code and models to be released for public use.

Abstract

Ultrasound plane wave imaging is a cutting-edge technique that enables high frame-rate imaging. However, one challenge associated with high frame-rate ultrasound imaging is the high noise associated with them, hindering their wider adoption. Therefore, the development of a denoising method becomes imperative to augment the quality of plane wave images. Drawing inspiration from Denoising Diffusion Probabilistic Models (DDPMs), our proposed solution aims to enhance plane wave image quality. Specifically, the method considers the distinction between low-angle and high-angle compounding plane waves as noise and effectively eliminates it by adapting a DDPM to beamformed radiofrequency (RF) data. The method underwent training using only 400 simulated images. In addition, our approach employs natural image segmentation masks as intensity maps for the generated images, resulting in accurate denoising for various anatomy shapes. The proposed method was assessed across simulation, phantom, and in vivo images. The results of the evaluations indicate that our approach not only enhances image quality on simulated data but also demonstrates effectiveness on phantom and in vivo data in terms of image quality. Comparative analysis with other methods underscores the superiority of our proposed method across various evaluation metrics. The source code and trained model will be released along with the dataset at: http://code.sonography.ai

Figures (9)

• Figure 1: Advantages and disadvantages of the generative models. Diffusion models generate high-quality and diverse results, albeit at a greater computational complexity.
• Figure 2: The QQ Plot corresponding to the difference between RF data reconstructed with one and 75 angles. A linear trend highlights a Gaussian distribution. The difference follows a Gaussian distribution in the middle, with deviations from this distribution in the first and last quantiles.
• Figure 3: An example of the forward diffusion process in 10 steps. The training occurs in the reverse process, while no learning takes place during the forward process.
• Figure 4: The proposed architecture, which is inspired by the U-Net.
• Figure 5: Samples of mask-simulated images used for the training procedure.