Abstract
Ultrasound (US) imaging is a fast and non-invasive imaging modality widely used for real-time clinical applications. However, it often suffers from poor visual quality due to speckle noise, blurring, and acquisition artifacts. While deep learning has successfully addressed these issues, the requirement for paired high-quality images for supervised training remains a significant hurdle. This paper investigates the applicability of unsupervised deep learning for US artifact removal using optimal transport-driven cycleGAN (OT-cycleGAN). Experimental results across various tasks, including deconvolution and speckle removal, confirm that this unsupervised approach provides results comparable to supervised learning without requiring matched reference data.
Introduction
Ultrasound imaging is essential for clinical diagnostics but is frequently compromised by artifacts such as speckle noise, resolution-limiting blur, and multi-line acquisition (MLA) artifacts. Deep learning models typically rely on supervised learning, which necessitates perfectly paired "noisy" and "clean" images. In practice, obtaining such pairs is difficult due to the dynamic nature of clinical scanning. This study addresses this limitation by leveraging unsupervised learning, specifically the OT-cycleGAN framework, to restore image quality from unpaired datasets.
Methodology
The core of the proposed method is the OT-cycleGAN, which maps probability distributions between the artifact-corrupted domain and the clean image domain. The architecture employs a U-Net as the generator and a PatchGAN as the discriminator. By incorporating Optimal Transport (OT) theory, the model ensures that the learned transformations are physically and mathematically consistent, avoiding the "hallucination" of features common in standard GANs.
Experimental Setup
The framework was evaluated using four datasets acquired with an E-CUBE 12R US system. These included linear array focused B-mode, linear array planewave B-mode, and phased array B-mode scans. The training utilized approximately 500 images per task, with "clean" distribution targets generated via established high-end algorithms like NLLR for despeckling.
Results and Discussion
The unsupervised model demonstrated significant performance gains across multiple artifact removal tasks. In deconvolution experiments, the method achieved Contrast Recovery (CR) improvements of 1.66 to 3.28 dB, matching or exceeding supervised benchmarks. Visual analysis shows that the unsupervised "DeepDeconv" method effectively recovers structural details in both phantom and in-vivo (trachea, right lobe) scans, closely approximating the target deconvolution results.
Furthermore, the model proved computationally efficient, with an inference time of 7.92 ms per frame, enabling real-time clinical deployment. It also showed remarkable versatility, handling various MLA acceleration factors with a single trained model.
Conclusion
This research pushes the boundaries of unsupervised learning in medical imaging by demonstrating that OT-cycleGAN can effectively remove complex ultrasound artifacts without paired training data. The results suggest that unsupervised frameworks can achieve clinical-grade restoration quality while offering the flexibility and speed required for real-time diagnostic applications.
Summary updated