Robust Medical Ultrasound Analysis Under Acoustic Artifacts

84 concepts83 papers2 proposals

Medical ultrasound is one of the most widely used diagnostic imaging modalities due to its safety, real-time capabilities, and cost-effectiveness. However, automated computer vision systems consistently struggle to interpret ultrasound images. Unlike Computed Tomography (CT) or Magnetic Resonance Imaging (MRI), which produce highly structured, high-contrast representations of tissue properties, ultrasound images are fundamentally shaped by the physics of acoustic wave propagation.

When sound waves travel through tissue, micro-structures cause them to scatter, resulting in constructive and destructive wave interference. This manifests as speckle noise—a granular, high-frequency texture that standard convolutional networks often mistake for legitimate tissue boundaries, as noted by Dimitris Perdios et al. in CNN-Based Image Reconstruction Method for Ultrafast Ultrasound Imaging. Furthermore, dense or highly reflective structures (such as bones or calcifications) absorb or reflect nearly all acoustic energy, creating acoustic shadows that completely obliterate downstream anatomical information, a challenge highlighted by Farhang Sahba et al. in A coarse-to-fine approach to prostate boundary segmentation in ultrasound images.

To build automated segmentation, tracking, and localization models that work reliably in clinical settings, researchers have moved away from generic computer vision architectures. Instead, modern approaches explicitly exploit the temporal, structural, and physical properties of ultrasound imaging to overcome these acoustic artifacts.

Leveraging Spatio-Temporal Continuity

While a single ultrasound frame may be heavily degraded by a transient artifact or a localized shadow, anatomical structures move continuously over time. In video sequences—such as echocardiograms or fetal scans—the underlying tissue follows strict physical trajectories, whereas speckle noise and acoustic shadows fluctuate or behave predictably relative to the probe's position.

By expanding the optimization space from isolated 2D frames to continuous spatio-temporal volumes, models can "fill in" missing visual data. For instance, María J. Ledesma‐Carbayo et al. demonstrated in Spatio-temporal nonrigid registration for ultrasound cardiac motion estimation that modeling cardiac deformation simultaneously across space and time using parametric B-splines prevents tracking algorithms from getting derailed by localized shadowing. Because the mathematical model enforces temporal smoothness, the system relies on surrounding frames to infer motion where the signal is temporarily lost.

Similarly, foundation models can exploit the cyclical nature of cardiac motion to build noise-resilient representations. In EchoFM: Foundation Model for Generalizable Echocardiogram Analysis, Sekeun Kim et al. utilized spatio-temporal consistent masking during self-supervised pre-training. By forcing the network to predict masked anatomical structures across cardiac cycles, the model learns to prioritize the underlying periodic motion of heart chambers over the chaotic, non-periodic variations of frame-level speckle noise.

Expanding Context via Multi-Scale Attention

Standard convolutional neural networks (CNNs) rely heavily on local texture gradients to identify boundaries. In ultrasound imaging, this local focus is a liability: a patch of speckle noise can look identical to a tissue interface, and blurry tumor boundaries provide weak local gradients. To resolve this ambiguity, models must look globally, using macro-structural context to infer the location of ambiguous local boundaries.

[Local Patch: Ambiguous Gradients] ---> CNN alone fails (Mistakes speckle for edge)
               |
               v
[Global Context / Macro-Structure] ---> Attention Mechanism guides the model
               |
               v
[Inferred Boundary Location] ---------> Robust Segmentation

To implement this intuition, Cheng Xue et al. designed the Global Guidance Network (GG-Net) in Global Guidance Network for Breast Lesion Segmentation in Ultrasound Images. GG-Net uses a dedicated block to capture long-range dependencies across both spatial and channel domains, ensuring that the network's local segmentation decisions are always constrained by the "big picture" geometry of the organ.

When targets vary significantly in size, single-scale global context is insufficient. Hao Zhang et al. addressed this in ACL-DUNet: A tumor segmentation method based on multiple attention and densely connected breast ultrasound images by introducing a triple-threat attention decoder. By simultaneously computing spatial, channel, and scale-specific attention, the network isolates irregular tumor boundaries from surrounding healthy tissue, effectively ignoring localized acoustic clutter by cross-referencing feature maps across multiple resolutions.

Baking Physical Priors into Neural Architectures

Rather than forcing a deep learning model to learn the physics of ultrasound entirely from data, researchers have found success by hardcoding physical relationships directly into the network's structure or training objectives.

A clear example of this is the relationship between bone surfaces and acoustic shadows. A shadow is not random noise; it is a deterministic physical consequence of an acoustic wave hitting a highly attenuating boundary. In Simultaneous Bone and Shadow Segmentation Network using Task Correspondence Consistency, Aimon Rahman et al. turned this destructive artifact into a supervisory signal. They developed a multi-task network that simultaneously segments bone surfaces and their corresponding shadows. By enforcing a "Task Correspondence Consistency" loss, the network learns that a bone boundary must have a shadow directly behind it. If the local bone signal is faint or obscured by noise, the presence of the shadow confirms the bone's location, transforming an artifact into a diagnostic asset.

Alternatively, the physics of ultrasound can be integrated via simulation to bridge the domain gap between imaging modalities. In LOTUS: Learning to Optimize Task-based US representations, Yordanka Velikova et al. introduced a differentiable ultrasound simulator into an end-to-end training pipeline. By simulating how sound waves interact with anatomical volumes derived from clean, easy-to-label CT scans, the framework explicitly teaches the segmentation network how speckle and clutter artifacts distort real-world ultrasound data.

Noise-Resilient Optimization and Hardness-Aware Learning

Standard segmentation losses, such as the Dice coefficient or cross-entropy, treat boundaries as sharp, binary pixel splits. However, ultrasound boundaries are inherently probabilistic and fuzzy due to acoustic blurring.

To capture this uncertainty mathematically, Md. Abdur Rahman et al. proposed a generative approach in Echo-DND: a dual noise diffusion model for robust and precise left ventricle segmentation in echocardiography. Instead of predicting a deterministic mask, their model (Echo-DND) utilizes a diffusion process that blends Gaussian and Bernoulli noise distributions. This dual-noise formulation is specifically tuned to model the statistical variance of echocardiographic boundaries, allowing the network to generate robust, stable segmentations of heart chambers despite heavy background noise.

Furthermore, because acoustic artifacts do not degrade an image uniformly, some regions are significantly more ambiguous than others. Standard data augmentation techniques can inadvertently destroy the subtle, low-contrast features that remain in these heavily obscured areas.

To protect these vulnerable regions, Huisi Wu introduced a semi-supervised framework in PH-Net: Semi-Supervised Breast Lesion Segmentation via Patch-Wise Hardness. PH-Net dynamically calculates a "hardness map" across image patches to identify regions heavily compromised by artifacts. Once these difficult patches are identified, the framework shields them from aggressive spatial augmentations and applies targeted contrastive learning. This ensures that the model actively learns to resolve ambiguous tissue interfaces rather than allowing the optimization process to be dominated by easily segmented, high-contrast regions.

Key Foundational Work

A coarse-to-fine approach to prostate boundary segmentation in ultrasound images (2005) — contributed 1 concept elements to this direction.
Toward Accurate Segmentation of the LV Myocardium and Chamber for Volumes Estimation in Gated SPECT Sequences (2004) — contributed 1 concept elements to this direction.
Spatio-temporal nonrigid registration for ultrasound cardiac motion estimation (2005) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_ultrasound_image_artifacts_87c4736e,
  title = {ultrasound image artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/ultrasound-image-artifacts/rdr_87c4736e88464c12bdab8a79dac1b759},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_ultrasound_image_artifacts_87c4736e_v1,
  title = {ultrasound image artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/ultrasound-image-artifacts/rdr_87c4736e88464c12bdab8a79dac1b759/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake