A Domain Adaptation Model for Carotid Ultrasound: Image Harmonization, Noise Reduction, and Impact on Cardiovascular Risk Markers

TL;DR

This work tackles distribution-shift issues in carotid ultrasound by introducing a GAN-based domain adaptation framework for image harmonization and noise reduction. The model employs a single generator and two discriminators, optimized with adversarial, content, and noise losses, including a Wasserstein-based noise term, to translate source-domain images to a target domain while preserving anatomy. It outperforms CycleGAN in feature-transfer metrics and maintains essential tissue structures, with significant improvements in risk-marker consistency in some experiments. The findings demonstrate that domain translation can enhance image quality and comparability across scanners, but underline the need to assess downstream risk-marker effects for each deployment due to domain-dependent changes in measurements like GSM.

Abstract

Deep learning has been used extensively for medical image analysis applications, assuming the training and test data adhere to the same probability distributions. However, a common challenge arises when dealing with medical images generated by different systems or even the same system with varying parameter settings. Such images often contain diverse textures and noise patterns, violating the assumption. Consequently, models trained on data from one machine or setting usually struggle to perform effectively on data from another. To address this issue in ultrasound images, we proposed a Generative Adversarial Network (GAN) based model in this paper. We formulated image harmonization and denoising tasks as an image-to-image translation task, wherein we modified the texture pattern and reduced noise in Carotid ultrasound images while keeping the image content (the anatomy) unchanged. The performance was evaluated using feature distribution and pixel-space similarity metrics. In addition, blood-to-tissue contrast and influence on computed risk markers (Gray scale median, GSM) were evaluated. The results showed that domain adaptation was achieved in both tasks (histogram correlation 0.920 and 0.844), as compared to no adaptation (0.890 and 0.707), and that the anatomy of the images was retained (structure similarity index measure of the arterial wall 0.71 and 0.80). In addition, the image noise level (contrast) did not change in the image harmonization task (-34.1 vs 35.2 dB) but was improved in the noise reduction task (-23.5 vs -46.7 dB). The model outperformed the CycleGAN in both tasks. Finally, the risk marker GSM increased by 7.6 (p<0.001) in task 1 but not in task 2. We conclude that domain translation models are powerful tools for ultrasound image improvement while retaining the underlying anatomy but that downstream calculations of risk markers may be affected.

Figures (7)

• Figure 1: a) Proposed model. $G$ denotes the generator network and two discriminators, $D_c$ and $D_n$, trained by adversarial losses $\mathcal{L}_{ac}$ and $\mathcal{L}_{an}$, respectively. $x \sim S$ denotes the image set from the source domain. $y \sim T$ represents the image set from the target domain. $G(x)$ denotes the image set generated by $G$. $L_c$ indicates the content loss calculated from the feature difference of the source and $G(x)$ image. $L_n$ represents the noise loss calculated from the feature difference between the target image and the early layers of the generator. b) a residual block.
• Figure 2: Examples of carotid ultrasound images used in the training of the domain adaptation task for the two experiments. a) show examples for experiment 1 (image harmonization) that we define as domains A and B, respectively. b) show examples for experiment 2 (noise reduction) that we define as domains C and D. In the lower panel, c) and d), regions of interests (ROIs) from the upper images are shown to zoom in on a homogeneous tissue segment (A1 and B1), carotid plaques (A2 and B2), Lumen/blood (C1 and D1), and arterial walls (C2 and D2).
• Figure 3: Example of results for experiment 1 (Image harmonization). a) and d) shows input images (Domain A). b) and e) shows the proposed model's generated images. c) and f) shows the corresponding generated images of the cycleGAN model.
• Figure 4: Summary of results. The upper panel is for experiment 1 (image harmonization), and the lower panel is for experiment 2 (noise reduction). The evaluation was done in four perspectives: 1) feature distributions, a) and d), 2) pixel space b) and e), and 3) image-based risk marker for cardiovascular disease (GSM) and 4) noise level (Contrast), c) and f). Feature distributions were compared using bhattacharyya distance (BD) and histogram correlation (HC), and pixel-space similarity was evaluated using the structure similarity index measure (SSIM). The risk marker and contrast measures are presented as the relative change between the corresponding input and domain-adapted output images.
• Figure 5: Illustration of ROIs used in Experiment 1 to assess translation performance (Lumen, Plaque and Wall/Adventitia). a) shows an example image of domain A, b) the generated image (domain B) and c) the corresponding structure similarity index measure (SSIM) map.