[Paper Review] DEMIX: Dual-Encoder Latent Masking Framework for Mixed Noise Reduction in Ultrasound Imaging

Soumee Guha, Scott T. Acton|arXiv (Cornell University)|Feb 6, 2026

Ultrasound Imaging and Elastography0 citations

TL;DR

DEMIX introduces a dual-encoder diffusion-inspired denoising framework that separately models additive and multiplicative noise in ultrasound images and fuses them with a masked gating mechanism, also accounting for PSF distortions; it outperforms state-of-the-art baselines on two ultrasound datasets and in downstream segmentation.

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.

Motivation & Objective

• Motivate robust ultrasound denoising under mixed noise (signal-dependent speckle and signal-independent Gaussian noise) and PSF-induced distortions.
• Propose a PSF-aware dual-encoder architecture that disentangles noise components in latent space.
• Introduce a masked gated fusion mechanism to combine additive and multiplicative noise representations.
• Incorporate PSF information into both encoders to adapt denoising to axial and lateral distortions.
• Demonstrate improved denoising quality and preserved structural details on two ultrasound datasets and in downstream segmentation.

Proposed method

• Propose a forward diffusion-like process modeling mixed noise as additive plus multiplicative perturbations convolved with a PSF.
• Introduce two encoders: one for additive noise and one for multiplicative noise, each encoding their respective noise characteristics.
• Use a gated fusion block to combine latent features from both encoders via a learnable gate and a masking strategy that ensures at least one encoder remains active.
• Embed noise schedules (alpha for multiplicative, beta for additive) and PSF parameters psi into the network to condition denoising across varying noise levels and PSF distortions.
• Employ a dual-encoder UNet backbone for the reverse (denoising) process, with the noise embedding alpha-beta conditioning all decoder stages.
• Optimize with a diffusion-style L1 loss between the denoised output and the clean image, augmented by a multiscale SSIM loss (MS-SSIM).

Experimental results

Research questions

• RQ1 Can a dual-encoder setup effectively disentangle signal-dependent speckle and signal-independent Gaussian noise in ultrasound images?
• RQ2 Does incorporating PSF (axial and lateral) information into each encoder improve restoration under PSF-induced distortions?
• RQ3 Does a masked gated fusion mechanism improve robustness and preservation of structural details during denoising of mixed-noise ultrasound images?
• RQ4 How does DEMIX perform relative to state-of-the-art denoising baselines on standard ultrasound datasets and in downstream segmentation tasks?

Key findings

• DEMIX consistently surpasses state-of-the-art baselines in PSNR and SSIM on two ultrasound datasets across varying noise and PSF conditions.
• The dual-encoder with masked fusion effectively disentangles additive and multiplicative noise components in the latent space.
• Encoding the full spectrum of PSF distortions into the PSF encoder enables adaptive restoration of deformities caused by axial and lateral PSF variations.
• The MS-SSIM-enhanced loss improves perceptual quality and maintains structural details in the reconstructed ultrasound images.
• A downstream segmentation task benefits from DEMIX-denoised inputs, indicating practical clinical utility.