[Paper Review] Pushing the Limit of Unsupervised Learning for Ultrasound Image Artifact Removal

Shujaat Khan, Jaeyoung Huh+1 more|arXiv (Cornell University)|Jun 26, 2020

Image and Signal Denoising Methods48 references4 citations

TL;DR

This paper proposes an unsupervised deep learning method for ultrasound image artifact removal using optimal transport-driven cycleGAN (OT-cycleGAN), enabling high-quality image reconstruction without paired training data. The approach achieves near-supervised performance across multiple artifact types—speckle noise, block artifacts, limited sampling, and plane-to-focused B-mode conversion—using a single, universally trained model with fast inference (7.92 ms/image).

ABSTRACT

Ultrasound (US) imaging is a fast and non-invasive imaging modality which is widely used for real-time clinical imaging applications without concerning about radiation hazard. Unfortunately, it often suffers from poor visual quality from various origins, such as speckle noises, blurring, multi-line acquisition (MLA), limited RF channels, small number of view angles for the case of plane wave imaging, etc. Classical methods to deal with these problems include image-domain signal processing approaches using various adaptive filtering and model-based approaches. Recently, deep learning approaches have been successfully used for ultrasound imaging field. However, one of the limitations of these approaches is that paired high quality images for supervised training are difficult to obtain in many practical applications. In this paper, inspired by the recent theory of unsupervised learning using optimal transport driven cycleGAN (OT-cycleGAN), we investigate applicability of unsupervised deep learning for US artifact removal problems without matched reference data. Experimental results for various tasks such as deconvolution, speckle removal, limited data artifact removal, etc. confirmed that our unsupervised learning method provides comparable results to supervised learning for many practical applications.

Motivation & Objective

• Address the lack of paired high-quality ultrasound images for supervised training in real-world clinical settings.
• Overcome limitations of classical filtering and model-based methods, which suffer from high computational cost and model inconsistency.
• Develop a robust, generalizable unsupervised deep learning framework for diverse ultrasound artifact removal tasks without requiring matched reference data.
• Validate that the learned image enhancement is not merely cosmetic but reflects real distributional improvements via optimal transport theory.
• Enable practical deployment in low-cost, portable, or fast imaging systems where high-end reference scans are infeasible.

Proposed method

• Adopt the optimal transport-driven cycleGAN (OT-cycleGAN) framework, which provides a theoretically grounded unsupervised learning mechanism for mapping noisy image distributions to clean image distributions.
• Train a single universal model on low-quality ultrasound images from one acquisition modality (e.g., plane wave or sub-sampled RF) and use high-quality images from different anatomical regions or imaging conditions as the target domain.
• Leverage cycle-consistency loss with optimal transport to ensure structural and distributional fidelity during image translation, minimizing artifact introduction.
• Use a single, one-time trained model for multiple tasks—deconvolution, speckle removal, limited data artifact suppression—across varying acceleration factors (SLA to 6-MLA) and sub-sampling rates.
• Optimize inference speed through efficient network design, achieving an average reconstruction time of 7.92 ms per image, suitable for real-time applications.
• Validate generalization using in-vivo data and tissue-mimicking phantoms to assess structural preservation and accuracy under diverse imaging conditions.

Experimental results

Research questions

• RQ1 Can unsupervised deep learning using OT-cycleGAN achieve artifact removal performance comparable to supervised methods without paired training data?
• RQ2 Does the OT-cycleGAN framework ensure that image quality improvements are not just visual artifacts but represent real distributional shifts toward clean images?
• RQ3 Can a single, universally trained model effectively handle multiple ultrasound artifact types—speckle noise, block artifacts, limited sampling—across different imaging protocols?
• RQ4 How does the method perform in terms of quantitative metrics (e.g., GCNR) and visual quality across varying acceleration factors and sub-sampling rates?
• RQ5 Is the method robust to anatomical variability and capable of preserving fine structures without introducing false features?

Key findings

• The proposed unsupervised OT-cycleGAN method achieved an average GCNR gain of 0.1372 across all MLA acceleration factors (SLA to 6-MLA), significantly improving image contrast-to-noise ratio.
• For sub-sampled RF data, the unsupervised model achieved a CR (contrast-to-noise ratio) of 11.06 dB at 1× sub-sampling, approaching the supervised baseline of 9.80 dB at 1× sub-sampling.
• The method maintained consistent performance across all tested acceleration factors (SLA to 6-MLA) with no retraining, demonstrating strong generalization and robustness.
• Reconstruction time averaged 7.92 ms per image, enabling real-time applicability in clinical settings such as echocardiography.
• In phantom studies, the method accurately recovered structural details, though performance degraded slightly at high acceleration factors (e.g., 6-MLA), where block artifacts remained prominent.
• Visual and quantitative results confirmed that the enhancement was not cosmetic: the method preserved tissue structures and reduced both speckle and block artifacts effectively across diverse imaging conditions.