DEMIX: Dual-Encoder Latent Masking Framework for Mixed Noise Reduction in Ultrasound Imaging

TL;DR

DEMIX addresses mixed noise and PSF distortions in ultrasound imaging by introducing a diffusion-inspired dual-encoder denoising framework with latent masking and gated fusion. The model encodes additive and multiplicative noise separately and conditions restoration on a PSF embedding, allowing adaptive suppression of noise while preserving structural details. On two ultrasound datasets, DEMIX achieves superior PSNR/SSIM and improves downstream segmentation, demonstrating robustness to varying noise and PSF conditions with patch-efficient training. This approach offers a practical, PSF-aware solution for real-world ultrasound denoising and holds potential for other coherent-imaging modalities such as microscopy and astronomy.

Abstract

Ultrasound imaging is widely used in noninvasive medical diagnostics due to its efficiency, portability, and avoidance of ionizing radiation. However, its utility is limited by the quality of the signal. Signal-dependent speckle noise, signal-independent sensor noise, and non-uniform spatial blurring caused by the transducer and modeled by the point spread function (PSF) degrade the image quality. These degradations challenge conventional image restoration methods, which assume simplified noise models, and highlight the need for specialized algorithms capable of effectively reducing the degradations while preserving fine structural details. We propose DEMIX, a novel dual-encoder denoising framework with a masked gated fusion mechanism, for denoising ultrasound images degraded by mixed noise and further degraded by PSF-induced distortions. DEMIX is inspired by diffusion models and is characterized by a forward process and a deterministic reverse process. DEMIX adaptively assesses the different noise components, disentangles them in the latent space, and suppresses these components while compensating for PSF degradations. Extensive experiments on two ultrasound datasets, along with a downstream segmentation task, demonstrate that DEMIX consistently outperforms state-of-the-art baselines, achieving superior noise suppression and preserving structural details. The code will be made publicly available.

Figures (7)

• Figure 1: Noisy image generation: The original image is first multiplied with signal-dependent speckle, convolved with the PSF of the imaging system and finally added to signal-independent sensor noise.
• Figure 2: (a) The overall architecture is shown and $\alpha$ and $\beta$ are the multiplicative and additive schedules. The dual encoder architecture processes and extracts the additive and multiplicative noise features separately and combines them through the gated fusion block. The combined features are subsequently processed by the decoder layers, and the final denoised image is obtained. (b) The noise encoder combines the multiplicative and additive noise schedules ($\alpha$ and $\beta$, respectively) and embeds them together to create the final noise embedding vector $\alpha \beta$, which is combined with the bottleneck layer and all decoder layers of DEMIX. (c) The gated fusion block combines the multiplicative $\chi_{mul}$ and additive $\chi_{add}$ noise features, along with the sampled mask and fuses them to obtain a feature vector $\chi$ for further processing in the bottleneck layer and the decoder layers for efficient denoising performance.
• Figure 3: Qualitative comparisons of different denoising methods for $\alpha_t = 0.373, \beta_t = 0.127$ and $\sigma_x = 3, \sigma_y = 2.5$. Column 1: Ground truth images from ultrasound dataset (row 1) and the phantom dataset (row 2). Column 2: Noisy images. Columns 3 - 12: Denoised images by different algorithms. The results for the proposed method DEMIX are shown in the last column.
• Figure 4: (a) The MS-SSIM loss (shown for $\sigma_x = 2, \sigma_y = 1.5$) improves the PSNR and SSIM scores for both datasets for the proposed method DEMIX. (b) With MS-SSIM loss, the PSNR and SSIM scores (shown for $\sigma_x = 2, \sigma_y = 1.5$) for DEMIX are consistently higher than the ablated models without the gated fusion and the noise encoder.
• Figure 5: (a) $\sigma_x = 3, \sigma_y = 2.5$, $\alpha_t = 0.598$, $\beta_t = 0.201$ (b) For PSF parameters $\sigma_x = 2, \sigma_y = 1.5$, we show how the PSNR and SSIM scores vary for the different noise combinations for the two datasets. All other PSF distorted images follow similar trends.